Loading...
1/*
2 * Copyright (C) 2009 Red Hat, Inc.
3 *
4 * This work is licensed under the terms of the GNU GPL, version 2. See
5 * the COPYING file in the top-level directory.
6 */
7
8#include <linux/mm.h>
9#include <linux/sched.h>
10#include <linux/highmem.h>
11#include <linux/hugetlb.h>
12#include <linux/mmu_notifier.h>
13#include <linux/rmap.h>
14#include <linux/swap.h>
15#include <linux/shrinker.h>
16#include <linux/mm_inline.h>
17#include <linux/kthread.h>
18#include <linux/khugepaged.h>
19#include <linux/freezer.h>
20#include <linux/mman.h>
21#include <linux/pagemap.h>
22#include <linux/migrate.h>
23#include <linux/hashtable.h>
24
25#include <asm/tlb.h>
26#include <asm/pgalloc.h>
27#include "internal.h"
28
29/*
30 * By default transparent hugepage support is disabled in order that avoid
31 * to risk increase the memory footprint of applications without a guaranteed
32 * benefit. When transparent hugepage support is enabled, is for all mappings,
33 * and khugepaged scans all mappings.
34 * Defrag is invoked by khugepaged hugepage allocations and by page faults
35 * for all hugepage allocations.
36 */
37unsigned long transparent_hugepage_flags __read_mostly =
38#ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
39 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
40#endif
41#ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
42 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
43#endif
44 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_FLAG)|
45 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
46 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
47
48/* default scan 8*512 pte (or vmas) every 30 second */
49static unsigned int khugepaged_pages_to_scan __read_mostly = HPAGE_PMD_NR*8;
50static unsigned int khugepaged_pages_collapsed;
51static unsigned int khugepaged_full_scans;
52static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
53/* during fragmentation poll the hugepage allocator once every minute */
54static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
55static struct task_struct *khugepaged_thread __read_mostly;
56static DEFINE_MUTEX(khugepaged_mutex);
57static DEFINE_SPINLOCK(khugepaged_mm_lock);
58static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
59/*
60 * default collapse hugepages if there is at least one pte mapped like
61 * it would have happened if the vma was large enough during page
62 * fault.
63 */
64static unsigned int khugepaged_max_ptes_none __read_mostly = HPAGE_PMD_NR-1;
65
66static int khugepaged(void *none);
67static int khugepaged_slab_init(void);
68
69#define MM_SLOTS_HASH_BITS 10
70static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
71
72static struct kmem_cache *mm_slot_cache __read_mostly;
73
74/**
75 * struct mm_slot - hash lookup from mm to mm_slot
76 * @hash: hash collision list
77 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
78 * @mm: the mm that this information is valid for
79 */
80struct mm_slot {
81 struct hlist_node hash;
82 struct list_head mm_node;
83 struct mm_struct *mm;
84};
85
86/**
87 * struct khugepaged_scan - cursor for scanning
88 * @mm_head: the head of the mm list to scan
89 * @mm_slot: the current mm_slot we are scanning
90 * @address: the next address inside that to be scanned
91 *
92 * There is only the one khugepaged_scan instance of this cursor structure.
93 */
94struct khugepaged_scan {
95 struct list_head mm_head;
96 struct mm_slot *mm_slot;
97 unsigned long address;
98};
99static struct khugepaged_scan khugepaged_scan = {
100 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
101};
102
103
104static int set_recommended_min_free_kbytes(void)
105{
106 struct zone *zone;
107 int nr_zones = 0;
108 unsigned long recommended_min;
109
110 if (!khugepaged_enabled())
111 return 0;
112
113 for_each_populated_zone(zone)
114 nr_zones++;
115
116 /* Make sure at least 2 hugepages are free for MIGRATE_RESERVE */
117 recommended_min = pageblock_nr_pages * nr_zones * 2;
118
119 /*
120 * Make sure that on average at least two pageblocks are almost free
121 * of another type, one for a migratetype to fall back to and a
122 * second to avoid subsequent fallbacks of other types There are 3
123 * MIGRATE_TYPES we care about.
124 */
125 recommended_min += pageblock_nr_pages * nr_zones *
126 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
127
128 /* don't ever allow to reserve more than 5% of the lowmem */
129 recommended_min = min(recommended_min,
130 (unsigned long) nr_free_buffer_pages() / 20);
131 recommended_min <<= (PAGE_SHIFT-10);
132
133 if (recommended_min > min_free_kbytes) {
134 if (user_min_free_kbytes >= 0)
135 pr_info("raising min_free_kbytes from %d to %lu "
136 "to help transparent hugepage allocations\n",
137 min_free_kbytes, recommended_min);
138
139 min_free_kbytes = recommended_min;
140 }
141 setup_per_zone_wmarks();
142 return 0;
143}
144late_initcall(set_recommended_min_free_kbytes);
145
146static int start_khugepaged(void)
147{
148 int err = 0;
149 if (khugepaged_enabled()) {
150 if (!khugepaged_thread)
151 khugepaged_thread = kthread_run(khugepaged, NULL,
152 "khugepaged");
153 if (unlikely(IS_ERR(khugepaged_thread))) {
154 printk(KERN_ERR
155 "khugepaged: kthread_run(khugepaged) failed\n");
156 err = PTR_ERR(khugepaged_thread);
157 khugepaged_thread = NULL;
158 }
159
160 if (!list_empty(&khugepaged_scan.mm_head))
161 wake_up_interruptible(&khugepaged_wait);
162
163 set_recommended_min_free_kbytes();
164 } else if (khugepaged_thread) {
165 kthread_stop(khugepaged_thread);
166 khugepaged_thread = NULL;
167 }
168
169 return err;
170}
171
172static atomic_t huge_zero_refcount;
173static struct page *huge_zero_page __read_mostly;
174
175static inline bool is_huge_zero_page(struct page *page)
176{
177 return ACCESS_ONCE(huge_zero_page) == page;
178}
179
180static inline bool is_huge_zero_pmd(pmd_t pmd)
181{
182 return is_huge_zero_page(pmd_page(pmd));
183}
184
185static struct page *get_huge_zero_page(void)
186{
187 struct page *zero_page;
188retry:
189 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
190 return ACCESS_ONCE(huge_zero_page);
191
192 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
193 HPAGE_PMD_ORDER);
194 if (!zero_page) {
195 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
196 return NULL;
197 }
198 count_vm_event(THP_ZERO_PAGE_ALLOC);
199 preempt_disable();
200 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
201 preempt_enable();
202 __free_page(zero_page);
203 goto retry;
204 }
205
206 /* We take additional reference here. It will be put back by shrinker */
207 atomic_set(&huge_zero_refcount, 2);
208 preempt_enable();
209 return ACCESS_ONCE(huge_zero_page);
210}
211
212static void put_huge_zero_page(void)
213{
214 /*
215 * Counter should never go to zero here. Only shrinker can put
216 * last reference.
217 */
218 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
219}
220
221static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
222 struct shrink_control *sc)
223{
224 /* we can free zero page only if last reference remains */
225 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
226}
227
228static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
229 struct shrink_control *sc)
230{
231 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
232 struct page *zero_page = xchg(&huge_zero_page, NULL);
233 BUG_ON(zero_page == NULL);
234 __free_page(zero_page);
235 return HPAGE_PMD_NR;
236 }
237
238 return 0;
239}
240
241static struct shrinker huge_zero_page_shrinker = {
242 .count_objects = shrink_huge_zero_page_count,
243 .scan_objects = shrink_huge_zero_page_scan,
244 .seeks = DEFAULT_SEEKS,
245};
246
247#ifdef CONFIG_SYSFS
248
249static ssize_t double_flag_show(struct kobject *kobj,
250 struct kobj_attribute *attr, char *buf,
251 enum transparent_hugepage_flag enabled,
252 enum transparent_hugepage_flag req_madv)
253{
254 if (test_bit(enabled, &transparent_hugepage_flags)) {
255 VM_BUG_ON(test_bit(req_madv, &transparent_hugepage_flags));
256 return sprintf(buf, "[always] madvise never\n");
257 } else if (test_bit(req_madv, &transparent_hugepage_flags))
258 return sprintf(buf, "always [madvise] never\n");
259 else
260 return sprintf(buf, "always madvise [never]\n");
261}
262static ssize_t double_flag_store(struct kobject *kobj,
263 struct kobj_attribute *attr,
264 const char *buf, size_t count,
265 enum transparent_hugepage_flag enabled,
266 enum transparent_hugepage_flag req_madv)
267{
268 if (!memcmp("always", buf,
269 min(sizeof("always")-1, count))) {
270 set_bit(enabled, &transparent_hugepage_flags);
271 clear_bit(req_madv, &transparent_hugepage_flags);
272 } else if (!memcmp("madvise", buf,
273 min(sizeof("madvise")-1, count))) {
274 clear_bit(enabled, &transparent_hugepage_flags);
275 set_bit(req_madv, &transparent_hugepage_flags);
276 } else if (!memcmp("never", buf,
277 min(sizeof("never")-1, count))) {
278 clear_bit(enabled, &transparent_hugepage_flags);
279 clear_bit(req_madv, &transparent_hugepage_flags);
280 } else
281 return -EINVAL;
282
283 return count;
284}
285
286static ssize_t enabled_show(struct kobject *kobj,
287 struct kobj_attribute *attr, char *buf)
288{
289 return double_flag_show(kobj, attr, buf,
290 TRANSPARENT_HUGEPAGE_FLAG,
291 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
292}
293static ssize_t enabled_store(struct kobject *kobj,
294 struct kobj_attribute *attr,
295 const char *buf, size_t count)
296{
297 ssize_t ret;
298
299 ret = double_flag_store(kobj, attr, buf, count,
300 TRANSPARENT_HUGEPAGE_FLAG,
301 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
302
303 if (ret > 0) {
304 int err;
305
306 mutex_lock(&khugepaged_mutex);
307 err = start_khugepaged();
308 mutex_unlock(&khugepaged_mutex);
309
310 if (err)
311 ret = err;
312 }
313
314 return ret;
315}
316static struct kobj_attribute enabled_attr =
317 __ATTR(enabled, 0644, enabled_show, enabled_store);
318
319static ssize_t single_flag_show(struct kobject *kobj,
320 struct kobj_attribute *attr, char *buf,
321 enum transparent_hugepage_flag flag)
322{
323 return sprintf(buf, "%d\n",
324 !!test_bit(flag, &transparent_hugepage_flags));
325}
326
327static ssize_t single_flag_store(struct kobject *kobj,
328 struct kobj_attribute *attr,
329 const char *buf, size_t count,
330 enum transparent_hugepage_flag flag)
331{
332 unsigned long value;
333 int ret;
334
335 ret = kstrtoul(buf, 10, &value);
336 if (ret < 0)
337 return ret;
338 if (value > 1)
339 return -EINVAL;
340
341 if (value)
342 set_bit(flag, &transparent_hugepage_flags);
343 else
344 clear_bit(flag, &transparent_hugepage_flags);
345
346 return count;
347}
348
349/*
350 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
351 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
352 * memory just to allocate one more hugepage.
353 */
354static ssize_t defrag_show(struct kobject *kobj,
355 struct kobj_attribute *attr, char *buf)
356{
357 return double_flag_show(kobj, attr, buf,
358 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
359 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
360}
361static ssize_t defrag_store(struct kobject *kobj,
362 struct kobj_attribute *attr,
363 const char *buf, size_t count)
364{
365 return double_flag_store(kobj, attr, buf, count,
366 TRANSPARENT_HUGEPAGE_DEFRAG_FLAG,
367 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
368}
369static struct kobj_attribute defrag_attr =
370 __ATTR(defrag, 0644, defrag_show, defrag_store);
371
372static ssize_t use_zero_page_show(struct kobject *kobj,
373 struct kobj_attribute *attr, char *buf)
374{
375 return single_flag_show(kobj, attr, buf,
376 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
377}
378static ssize_t use_zero_page_store(struct kobject *kobj,
379 struct kobj_attribute *attr, const char *buf, size_t count)
380{
381 return single_flag_store(kobj, attr, buf, count,
382 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
383}
384static struct kobj_attribute use_zero_page_attr =
385 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
386#ifdef CONFIG_DEBUG_VM
387static ssize_t debug_cow_show(struct kobject *kobj,
388 struct kobj_attribute *attr, char *buf)
389{
390 return single_flag_show(kobj, attr, buf,
391 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
392}
393static ssize_t debug_cow_store(struct kobject *kobj,
394 struct kobj_attribute *attr,
395 const char *buf, size_t count)
396{
397 return single_flag_store(kobj, attr, buf, count,
398 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
399}
400static struct kobj_attribute debug_cow_attr =
401 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
402#endif /* CONFIG_DEBUG_VM */
403
404static struct attribute *hugepage_attr[] = {
405 &enabled_attr.attr,
406 &defrag_attr.attr,
407 &use_zero_page_attr.attr,
408#ifdef CONFIG_DEBUG_VM
409 &debug_cow_attr.attr,
410#endif
411 NULL,
412};
413
414static struct attribute_group hugepage_attr_group = {
415 .attrs = hugepage_attr,
416};
417
418static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
419 struct kobj_attribute *attr,
420 char *buf)
421{
422 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
423}
424
425static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
426 struct kobj_attribute *attr,
427 const char *buf, size_t count)
428{
429 unsigned long msecs;
430 int err;
431
432 err = kstrtoul(buf, 10, &msecs);
433 if (err || msecs > UINT_MAX)
434 return -EINVAL;
435
436 khugepaged_scan_sleep_millisecs = msecs;
437 wake_up_interruptible(&khugepaged_wait);
438
439 return count;
440}
441static struct kobj_attribute scan_sleep_millisecs_attr =
442 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
443 scan_sleep_millisecs_store);
444
445static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
446 struct kobj_attribute *attr,
447 char *buf)
448{
449 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
450}
451
452static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
453 struct kobj_attribute *attr,
454 const char *buf, size_t count)
455{
456 unsigned long msecs;
457 int err;
458
459 err = kstrtoul(buf, 10, &msecs);
460 if (err || msecs > UINT_MAX)
461 return -EINVAL;
462
463 khugepaged_alloc_sleep_millisecs = msecs;
464 wake_up_interruptible(&khugepaged_wait);
465
466 return count;
467}
468static struct kobj_attribute alloc_sleep_millisecs_attr =
469 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
470 alloc_sleep_millisecs_store);
471
472static ssize_t pages_to_scan_show(struct kobject *kobj,
473 struct kobj_attribute *attr,
474 char *buf)
475{
476 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
477}
478static ssize_t pages_to_scan_store(struct kobject *kobj,
479 struct kobj_attribute *attr,
480 const char *buf, size_t count)
481{
482 int err;
483 unsigned long pages;
484
485 err = kstrtoul(buf, 10, &pages);
486 if (err || !pages || pages > UINT_MAX)
487 return -EINVAL;
488
489 khugepaged_pages_to_scan = pages;
490
491 return count;
492}
493static struct kobj_attribute pages_to_scan_attr =
494 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
495 pages_to_scan_store);
496
497static ssize_t pages_collapsed_show(struct kobject *kobj,
498 struct kobj_attribute *attr,
499 char *buf)
500{
501 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
502}
503static struct kobj_attribute pages_collapsed_attr =
504 __ATTR_RO(pages_collapsed);
505
506static ssize_t full_scans_show(struct kobject *kobj,
507 struct kobj_attribute *attr,
508 char *buf)
509{
510 return sprintf(buf, "%u\n", khugepaged_full_scans);
511}
512static struct kobj_attribute full_scans_attr =
513 __ATTR_RO(full_scans);
514
515static ssize_t khugepaged_defrag_show(struct kobject *kobj,
516 struct kobj_attribute *attr, char *buf)
517{
518 return single_flag_show(kobj, attr, buf,
519 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
520}
521static ssize_t khugepaged_defrag_store(struct kobject *kobj,
522 struct kobj_attribute *attr,
523 const char *buf, size_t count)
524{
525 return single_flag_store(kobj, attr, buf, count,
526 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
527}
528static struct kobj_attribute khugepaged_defrag_attr =
529 __ATTR(defrag, 0644, khugepaged_defrag_show,
530 khugepaged_defrag_store);
531
532/*
533 * max_ptes_none controls if khugepaged should collapse hugepages over
534 * any unmapped ptes in turn potentially increasing the memory
535 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
536 * reduce the available free memory in the system as it
537 * runs. Increasing max_ptes_none will instead potentially reduce the
538 * free memory in the system during the khugepaged scan.
539 */
540static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
541 struct kobj_attribute *attr,
542 char *buf)
543{
544 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
545}
546static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
547 struct kobj_attribute *attr,
548 const char *buf, size_t count)
549{
550 int err;
551 unsigned long max_ptes_none;
552
553 err = kstrtoul(buf, 10, &max_ptes_none);
554 if (err || max_ptes_none > HPAGE_PMD_NR-1)
555 return -EINVAL;
556
557 khugepaged_max_ptes_none = max_ptes_none;
558
559 return count;
560}
561static struct kobj_attribute khugepaged_max_ptes_none_attr =
562 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
563 khugepaged_max_ptes_none_store);
564
565static struct attribute *khugepaged_attr[] = {
566 &khugepaged_defrag_attr.attr,
567 &khugepaged_max_ptes_none_attr.attr,
568 &pages_to_scan_attr.attr,
569 &pages_collapsed_attr.attr,
570 &full_scans_attr.attr,
571 &scan_sleep_millisecs_attr.attr,
572 &alloc_sleep_millisecs_attr.attr,
573 NULL,
574};
575
576static struct attribute_group khugepaged_attr_group = {
577 .attrs = khugepaged_attr,
578 .name = "khugepaged",
579};
580
581static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
582{
583 int err;
584
585 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
586 if (unlikely(!*hugepage_kobj)) {
587 printk(KERN_ERR "hugepage: failed to create transparent hugepage kobject\n");
588 return -ENOMEM;
589 }
590
591 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
592 if (err) {
593 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n");
594 goto delete_obj;
595 }
596
597 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
598 if (err) {
599 printk(KERN_ERR "hugepage: failed to register transparent hugepage group\n");
600 goto remove_hp_group;
601 }
602
603 return 0;
604
605remove_hp_group:
606 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
607delete_obj:
608 kobject_put(*hugepage_kobj);
609 return err;
610}
611
612static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
613{
614 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
615 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
616 kobject_put(hugepage_kobj);
617}
618#else
619static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
620{
621 return 0;
622}
623
624static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
625{
626}
627#endif /* CONFIG_SYSFS */
628
629static int __init hugepage_init(void)
630{
631 int err;
632 struct kobject *hugepage_kobj;
633
634 if (!has_transparent_hugepage()) {
635 transparent_hugepage_flags = 0;
636 return -EINVAL;
637 }
638
639 err = hugepage_init_sysfs(&hugepage_kobj);
640 if (err)
641 return err;
642
643 err = khugepaged_slab_init();
644 if (err)
645 goto out;
646
647 register_shrinker(&huge_zero_page_shrinker);
648
649 /*
650 * By default disable transparent hugepages on smaller systems,
651 * where the extra memory used could hurt more than TLB overhead
652 * is likely to save. The admin can still enable it through /sys.
653 */
654 if (totalram_pages < (512 << (20 - PAGE_SHIFT)))
655 transparent_hugepage_flags = 0;
656
657 start_khugepaged();
658
659 return 0;
660out:
661 hugepage_exit_sysfs(hugepage_kobj);
662 return err;
663}
664subsys_initcall(hugepage_init);
665
666static int __init setup_transparent_hugepage(char *str)
667{
668 int ret = 0;
669 if (!str)
670 goto out;
671 if (!strcmp(str, "always")) {
672 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
673 &transparent_hugepage_flags);
674 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
675 &transparent_hugepage_flags);
676 ret = 1;
677 } else if (!strcmp(str, "madvise")) {
678 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
679 &transparent_hugepage_flags);
680 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
681 &transparent_hugepage_flags);
682 ret = 1;
683 } else if (!strcmp(str, "never")) {
684 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
685 &transparent_hugepage_flags);
686 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
687 &transparent_hugepage_flags);
688 ret = 1;
689 }
690out:
691 if (!ret)
692 printk(KERN_WARNING
693 "transparent_hugepage= cannot parse, ignored\n");
694 return ret;
695}
696__setup("transparent_hugepage=", setup_transparent_hugepage);
697
698pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
699{
700 if (likely(vma->vm_flags & VM_WRITE))
701 pmd = pmd_mkwrite(pmd);
702 return pmd;
703}
704
705static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
706{
707 pmd_t entry;
708 entry = mk_pmd(page, prot);
709 entry = pmd_mkhuge(entry);
710 return entry;
711}
712
713static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
714 struct vm_area_struct *vma,
715 unsigned long haddr, pmd_t *pmd,
716 struct page *page)
717{
718 pgtable_t pgtable;
719 spinlock_t *ptl;
720
721 VM_BUG_ON_PAGE(!PageCompound(page), page);
722 pgtable = pte_alloc_one(mm, haddr);
723 if (unlikely(!pgtable))
724 return VM_FAULT_OOM;
725
726 clear_huge_page(page, haddr, HPAGE_PMD_NR);
727 /*
728 * The memory barrier inside __SetPageUptodate makes sure that
729 * clear_huge_page writes become visible before the set_pmd_at()
730 * write.
731 */
732 __SetPageUptodate(page);
733
734 ptl = pmd_lock(mm, pmd);
735 if (unlikely(!pmd_none(*pmd))) {
736 spin_unlock(ptl);
737 mem_cgroup_uncharge_page(page);
738 put_page(page);
739 pte_free(mm, pgtable);
740 } else {
741 pmd_t entry;
742 entry = mk_huge_pmd(page, vma->vm_page_prot);
743 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
744 page_add_new_anon_rmap(page, vma, haddr);
745 pgtable_trans_huge_deposit(mm, pmd, pgtable);
746 set_pmd_at(mm, haddr, pmd, entry);
747 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
748 atomic_long_inc(&mm->nr_ptes);
749 spin_unlock(ptl);
750 }
751
752 return 0;
753}
754
755static inline gfp_t alloc_hugepage_gfpmask(int defrag, gfp_t extra_gfp)
756{
757 return (GFP_TRANSHUGE & ~(defrag ? 0 : __GFP_WAIT)) | extra_gfp;
758}
759
760static inline struct page *alloc_hugepage_vma(int defrag,
761 struct vm_area_struct *vma,
762 unsigned long haddr, int nd,
763 gfp_t extra_gfp)
764{
765 return alloc_pages_vma(alloc_hugepage_gfpmask(defrag, extra_gfp),
766 HPAGE_PMD_ORDER, vma, haddr, nd);
767}
768
769/* Caller must hold page table lock. */
770static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
771 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
772 struct page *zero_page)
773{
774 pmd_t entry;
775 if (!pmd_none(*pmd))
776 return false;
777 entry = mk_pmd(zero_page, vma->vm_page_prot);
778 entry = pmd_wrprotect(entry);
779 entry = pmd_mkhuge(entry);
780 pgtable_trans_huge_deposit(mm, pmd, pgtable);
781 set_pmd_at(mm, haddr, pmd, entry);
782 atomic_long_inc(&mm->nr_ptes);
783 return true;
784}
785
786int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
787 unsigned long address, pmd_t *pmd,
788 unsigned int flags)
789{
790 struct page *page;
791 unsigned long haddr = address & HPAGE_PMD_MASK;
792
793 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
794 return VM_FAULT_FALLBACK;
795 if (unlikely(anon_vma_prepare(vma)))
796 return VM_FAULT_OOM;
797 if (unlikely(khugepaged_enter(vma)))
798 return VM_FAULT_OOM;
799 if (!(flags & FAULT_FLAG_WRITE) &&
800 transparent_hugepage_use_zero_page()) {
801 spinlock_t *ptl;
802 pgtable_t pgtable;
803 struct page *zero_page;
804 bool set;
805 pgtable = pte_alloc_one(mm, haddr);
806 if (unlikely(!pgtable))
807 return VM_FAULT_OOM;
808 zero_page = get_huge_zero_page();
809 if (unlikely(!zero_page)) {
810 pte_free(mm, pgtable);
811 count_vm_event(THP_FAULT_FALLBACK);
812 return VM_FAULT_FALLBACK;
813 }
814 ptl = pmd_lock(mm, pmd);
815 set = set_huge_zero_page(pgtable, mm, vma, haddr, pmd,
816 zero_page);
817 spin_unlock(ptl);
818 if (!set) {
819 pte_free(mm, pgtable);
820 put_huge_zero_page();
821 }
822 return 0;
823 }
824 page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
825 vma, haddr, numa_node_id(), 0);
826 if (unlikely(!page)) {
827 count_vm_event(THP_FAULT_FALLBACK);
828 return VM_FAULT_FALLBACK;
829 }
830 if (unlikely(mem_cgroup_charge_anon(page, mm, GFP_KERNEL))) {
831 put_page(page);
832 count_vm_event(THP_FAULT_FALLBACK);
833 return VM_FAULT_FALLBACK;
834 }
835 if (unlikely(__do_huge_pmd_anonymous_page(mm, vma, haddr, pmd, page))) {
836 mem_cgroup_uncharge_page(page);
837 put_page(page);
838 count_vm_event(THP_FAULT_FALLBACK);
839 return VM_FAULT_FALLBACK;
840 }
841
842 count_vm_event(THP_FAULT_ALLOC);
843 return 0;
844}
845
846int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
847 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
848 struct vm_area_struct *vma)
849{
850 spinlock_t *dst_ptl, *src_ptl;
851 struct page *src_page;
852 pmd_t pmd;
853 pgtable_t pgtable;
854 int ret;
855
856 ret = -ENOMEM;
857 pgtable = pte_alloc_one(dst_mm, addr);
858 if (unlikely(!pgtable))
859 goto out;
860
861 dst_ptl = pmd_lock(dst_mm, dst_pmd);
862 src_ptl = pmd_lockptr(src_mm, src_pmd);
863 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
864
865 ret = -EAGAIN;
866 pmd = *src_pmd;
867 if (unlikely(!pmd_trans_huge(pmd))) {
868 pte_free(dst_mm, pgtable);
869 goto out_unlock;
870 }
871 /*
872 * When page table lock is held, the huge zero pmd should not be
873 * under splitting since we don't split the page itself, only pmd to
874 * a page table.
875 */
876 if (is_huge_zero_pmd(pmd)) {
877 struct page *zero_page;
878 bool set;
879 /*
880 * get_huge_zero_page() will never allocate a new page here,
881 * since we already have a zero page to copy. It just takes a
882 * reference.
883 */
884 zero_page = get_huge_zero_page();
885 set = set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
886 zero_page);
887 BUG_ON(!set); /* unexpected !pmd_none(dst_pmd) */
888 ret = 0;
889 goto out_unlock;
890 }
891
892 if (unlikely(pmd_trans_splitting(pmd))) {
893 /* split huge page running from under us */
894 spin_unlock(src_ptl);
895 spin_unlock(dst_ptl);
896 pte_free(dst_mm, pgtable);
897
898 wait_split_huge_page(vma->anon_vma, src_pmd); /* src_vma */
899 goto out;
900 }
901 src_page = pmd_page(pmd);
902 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
903 get_page(src_page);
904 page_dup_rmap(src_page);
905 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
906
907 pmdp_set_wrprotect(src_mm, addr, src_pmd);
908 pmd = pmd_mkold(pmd_wrprotect(pmd));
909 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
910 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
911 atomic_long_inc(&dst_mm->nr_ptes);
912
913 ret = 0;
914out_unlock:
915 spin_unlock(src_ptl);
916 spin_unlock(dst_ptl);
917out:
918 return ret;
919}
920
921void huge_pmd_set_accessed(struct mm_struct *mm,
922 struct vm_area_struct *vma,
923 unsigned long address,
924 pmd_t *pmd, pmd_t orig_pmd,
925 int dirty)
926{
927 spinlock_t *ptl;
928 pmd_t entry;
929 unsigned long haddr;
930
931 ptl = pmd_lock(mm, pmd);
932 if (unlikely(!pmd_same(*pmd, orig_pmd)))
933 goto unlock;
934
935 entry = pmd_mkyoung(orig_pmd);
936 haddr = address & HPAGE_PMD_MASK;
937 if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
938 update_mmu_cache_pmd(vma, address, pmd);
939
940unlock:
941 spin_unlock(ptl);
942}
943
944static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
945 struct vm_area_struct *vma,
946 unsigned long address,
947 pmd_t *pmd, pmd_t orig_pmd,
948 struct page *page,
949 unsigned long haddr)
950{
951 spinlock_t *ptl;
952 pgtable_t pgtable;
953 pmd_t _pmd;
954 int ret = 0, i;
955 struct page **pages;
956 unsigned long mmun_start; /* For mmu_notifiers */
957 unsigned long mmun_end; /* For mmu_notifiers */
958
959 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
960 GFP_KERNEL);
961 if (unlikely(!pages)) {
962 ret |= VM_FAULT_OOM;
963 goto out;
964 }
965
966 for (i = 0; i < HPAGE_PMD_NR; i++) {
967 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
968 __GFP_OTHER_NODE,
969 vma, address, page_to_nid(page));
970 if (unlikely(!pages[i] ||
971 mem_cgroup_charge_anon(pages[i], mm,
972 GFP_KERNEL))) {
973 if (pages[i])
974 put_page(pages[i]);
975 mem_cgroup_uncharge_start();
976 while (--i >= 0) {
977 mem_cgroup_uncharge_page(pages[i]);
978 put_page(pages[i]);
979 }
980 mem_cgroup_uncharge_end();
981 kfree(pages);
982 ret |= VM_FAULT_OOM;
983 goto out;
984 }
985 }
986
987 for (i = 0; i < HPAGE_PMD_NR; i++) {
988 copy_user_highpage(pages[i], page + i,
989 haddr + PAGE_SIZE * i, vma);
990 __SetPageUptodate(pages[i]);
991 cond_resched();
992 }
993
994 mmun_start = haddr;
995 mmun_end = haddr + HPAGE_PMD_SIZE;
996 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
997
998 ptl = pmd_lock(mm, pmd);
999 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1000 goto out_free_pages;
1001 VM_BUG_ON_PAGE(!PageHead(page), page);
1002
1003 pmdp_clear_flush(vma, haddr, pmd);
1004 /* leave pmd empty until pte is filled */
1005
1006 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1007 pmd_populate(mm, &_pmd, pgtable);
1008
1009 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1010 pte_t *pte, entry;
1011 entry = mk_pte(pages[i], vma->vm_page_prot);
1012 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1013 page_add_new_anon_rmap(pages[i], vma, haddr);
1014 pte = pte_offset_map(&_pmd, haddr);
1015 VM_BUG_ON(!pte_none(*pte));
1016 set_pte_at(mm, haddr, pte, entry);
1017 pte_unmap(pte);
1018 }
1019 kfree(pages);
1020
1021 smp_wmb(); /* make pte visible before pmd */
1022 pmd_populate(mm, pmd, pgtable);
1023 page_remove_rmap(page);
1024 spin_unlock(ptl);
1025
1026 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1027
1028 ret |= VM_FAULT_WRITE;
1029 put_page(page);
1030
1031out:
1032 return ret;
1033
1034out_free_pages:
1035 spin_unlock(ptl);
1036 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1037 mem_cgroup_uncharge_start();
1038 for (i = 0; i < HPAGE_PMD_NR; i++) {
1039 mem_cgroup_uncharge_page(pages[i]);
1040 put_page(pages[i]);
1041 }
1042 mem_cgroup_uncharge_end();
1043 kfree(pages);
1044 goto out;
1045}
1046
1047int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1048 unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1049{
1050 spinlock_t *ptl;
1051 int ret = 0;
1052 struct page *page = NULL, *new_page;
1053 unsigned long haddr;
1054 unsigned long mmun_start; /* For mmu_notifiers */
1055 unsigned long mmun_end; /* For mmu_notifiers */
1056
1057 ptl = pmd_lockptr(mm, pmd);
1058 VM_BUG_ON(!vma->anon_vma);
1059 haddr = address & HPAGE_PMD_MASK;
1060 if (is_huge_zero_pmd(orig_pmd))
1061 goto alloc;
1062 spin_lock(ptl);
1063 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1064 goto out_unlock;
1065
1066 page = pmd_page(orig_pmd);
1067 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1068 if (page_mapcount(page) == 1) {
1069 pmd_t entry;
1070 entry = pmd_mkyoung(orig_pmd);
1071 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1072 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1))
1073 update_mmu_cache_pmd(vma, address, pmd);
1074 ret |= VM_FAULT_WRITE;
1075 goto out_unlock;
1076 }
1077 get_page(page);
1078 spin_unlock(ptl);
1079alloc:
1080 if (transparent_hugepage_enabled(vma) &&
1081 !transparent_hugepage_debug_cow())
1082 new_page = alloc_hugepage_vma(transparent_hugepage_defrag(vma),
1083 vma, haddr, numa_node_id(), 0);
1084 else
1085 new_page = NULL;
1086
1087 if (unlikely(!new_page)) {
1088 if (!page) {
1089 split_huge_page_pmd(vma, address, pmd);
1090 ret |= VM_FAULT_FALLBACK;
1091 } else {
1092 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1093 pmd, orig_pmd, page, haddr);
1094 if (ret & VM_FAULT_OOM) {
1095 split_huge_page(page);
1096 ret |= VM_FAULT_FALLBACK;
1097 }
1098 put_page(page);
1099 }
1100 count_vm_event(THP_FAULT_FALLBACK);
1101 goto out;
1102 }
1103
1104 if (unlikely(mem_cgroup_charge_anon(new_page, mm, GFP_KERNEL))) {
1105 put_page(new_page);
1106 if (page) {
1107 split_huge_page(page);
1108 put_page(page);
1109 } else
1110 split_huge_page_pmd(vma, address, pmd);
1111 ret |= VM_FAULT_FALLBACK;
1112 count_vm_event(THP_FAULT_FALLBACK);
1113 goto out;
1114 }
1115
1116 count_vm_event(THP_FAULT_ALLOC);
1117
1118 if (!page)
1119 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1120 else
1121 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1122 __SetPageUptodate(new_page);
1123
1124 mmun_start = haddr;
1125 mmun_end = haddr + HPAGE_PMD_SIZE;
1126 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1127
1128 spin_lock(ptl);
1129 if (page)
1130 put_page(page);
1131 if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1132 spin_unlock(ptl);
1133 mem_cgroup_uncharge_page(new_page);
1134 put_page(new_page);
1135 goto out_mn;
1136 } else {
1137 pmd_t entry;
1138 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1139 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1140 pmdp_clear_flush(vma, haddr, pmd);
1141 page_add_new_anon_rmap(new_page, vma, haddr);
1142 set_pmd_at(mm, haddr, pmd, entry);
1143 update_mmu_cache_pmd(vma, address, pmd);
1144 if (!page) {
1145 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1146 put_huge_zero_page();
1147 } else {
1148 VM_BUG_ON_PAGE(!PageHead(page), page);
1149 page_remove_rmap(page);
1150 put_page(page);
1151 }
1152 ret |= VM_FAULT_WRITE;
1153 }
1154 spin_unlock(ptl);
1155out_mn:
1156 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1157out:
1158 return ret;
1159out_unlock:
1160 spin_unlock(ptl);
1161 return ret;
1162}
1163
1164struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1165 unsigned long addr,
1166 pmd_t *pmd,
1167 unsigned int flags)
1168{
1169 struct mm_struct *mm = vma->vm_mm;
1170 struct page *page = NULL;
1171
1172 assert_spin_locked(pmd_lockptr(mm, pmd));
1173
1174 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1175 goto out;
1176
1177 /* Avoid dumping huge zero page */
1178 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1179 return ERR_PTR(-EFAULT);
1180
1181 /* Full NUMA hinting faults to serialise migration in fault paths */
1182 if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1183 goto out;
1184
1185 page = pmd_page(*pmd);
1186 VM_BUG_ON_PAGE(!PageHead(page), page);
1187 if (flags & FOLL_TOUCH) {
1188 pmd_t _pmd;
1189 /*
1190 * We should set the dirty bit only for FOLL_WRITE but
1191 * for now the dirty bit in the pmd is meaningless.
1192 * And if the dirty bit will become meaningful and
1193 * we'll only set it with FOLL_WRITE, an atomic
1194 * set_bit will be required on the pmd to set the
1195 * young bit, instead of the current set_pmd_at.
1196 */
1197 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1198 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1199 pmd, _pmd, 1))
1200 update_mmu_cache_pmd(vma, addr, pmd);
1201 }
1202 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1203 if (page->mapping && trylock_page(page)) {
1204 lru_add_drain();
1205 if (page->mapping)
1206 mlock_vma_page(page);
1207 unlock_page(page);
1208 }
1209 }
1210 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1211 VM_BUG_ON_PAGE(!PageCompound(page), page);
1212 if (flags & FOLL_GET)
1213 get_page_foll(page);
1214
1215out:
1216 return page;
1217}
1218
1219/* NUMA hinting page fault entry point for trans huge pmds */
1220int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1221 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1222{
1223 spinlock_t *ptl;
1224 struct anon_vma *anon_vma = NULL;
1225 struct page *page;
1226 unsigned long haddr = addr & HPAGE_PMD_MASK;
1227 int page_nid = -1, this_nid = numa_node_id();
1228 int target_nid, last_cpupid = -1;
1229 bool page_locked;
1230 bool migrated = false;
1231 int flags = 0;
1232
1233 ptl = pmd_lock(mm, pmdp);
1234 if (unlikely(!pmd_same(pmd, *pmdp)))
1235 goto out_unlock;
1236
1237 /*
1238 * If there are potential migrations, wait for completion and retry
1239 * without disrupting NUMA hinting information. Do not relock and
1240 * check_same as the page may no longer be mapped.
1241 */
1242 if (unlikely(pmd_trans_migrating(*pmdp))) {
1243 spin_unlock(ptl);
1244 wait_migrate_huge_page(vma->anon_vma, pmdp);
1245 goto out;
1246 }
1247
1248 page = pmd_page(pmd);
1249 BUG_ON(is_huge_zero_page(page));
1250 page_nid = page_to_nid(page);
1251 last_cpupid = page_cpupid_last(page);
1252 count_vm_numa_event(NUMA_HINT_FAULTS);
1253 if (page_nid == this_nid) {
1254 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1255 flags |= TNF_FAULT_LOCAL;
1256 }
1257
1258 /*
1259 * Avoid grouping on DSO/COW pages in specific and RO pages
1260 * in general, RO pages shouldn't hurt as much anyway since
1261 * they can be in shared cache state.
1262 */
1263 if (!pmd_write(pmd))
1264 flags |= TNF_NO_GROUP;
1265
1266 /*
1267 * Acquire the page lock to serialise THP migrations but avoid dropping
1268 * page_table_lock if at all possible
1269 */
1270 page_locked = trylock_page(page);
1271 target_nid = mpol_misplaced(page, vma, haddr);
1272 if (target_nid == -1) {
1273 /* If the page was locked, there are no parallel migrations */
1274 if (page_locked)
1275 goto clear_pmdnuma;
1276 }
1277
1278 /* Migration could have started since the pmd_trans_migrating check */
1279 if (!page_locked) {
1280 spin_unlock(ptl);
1281 wait_on_page_locked(page);
1282 page_nid = -1;
1283 goto out;
1284 }
1285
1286 /*
1287 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1288 * to serialises splits
1289 */
1290 get_page(page);
1291 spin_unlock(ptl);
1292 anon_vma = page_lock_anon_vma_read(page);
1293
1294 /* Confirm the PMD did not change while page_table_lock was released */
1295 spin_lock(ptl);
1296 if (unlikely(!pmd_same(pmd, *pmdp))) {
1297 unlock_page(page);
1298 put_page(page);
1299 page_nid = -1;
1300 goto out_unlock;
1301 }
1302
1303 /* Bail if we fail to protect against THP splits for any reason */
1304 if (unlikely(!anon_vma)) {
1305 put_page(page);
1306 page_nid = -1;
1307 goto clear_pmdnuma;
1308 }
1309
1310 /*
1311 * Migrate the THP to the requested node, returns with page unlocked
1312 * and pmd_numa cleared.
1313 */
1314 spin_unlock(ptl);
1315 migrated = migrate_misplaced_transhuge_page(mm, vma,
1316 pmdp, pmd, addr, page, target_nid);
1317 if (migrated) {
1318 flags |= TNF_MIGRATED;
1319 page_nid = target_nid;
1320 }
1321
1322 goto out;
1323clear_pmdnuma:
1324 BUG_ON(!PageLocked(page));
1325 pmd = pmd_mknonnuma(pmd);
1326 set_pmd_at(mm, haddr, pmdp, pmd);
1327 VM_BUG_ON(pmd_numa(*pmdp));
1328 update_mmu_cache_pmd(vma, addr, pmdp);
1329 unlock_page(page);
1330out_unlock:
1331 spin_unlock(ptl);
1332
1333out:
1334 if (anon_vma)
1335 page_unlock_anon_vma_read(anon_vma);
1336
1337 if (page_nid != -1)
1338 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1339
1340 return 0;
1341}
1342
1343int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1344 pmd_t *pmd, unsigned long addr)
1345{
1346 spinlock_t *ptl;
1347 int ret = 0;
1348
1349 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1350 struct page *page;
1351 pgtable_t pgtable;
1352 pmd_t orig_pmd;
1353 /*
1354 * For architectures like ppc64 we look at deposited pgtable
1355 * when calling pmdp_get_and_clear. So do the
1356 * pgtable_trans_huge_withdraw after finishing pmdp related
1357 * operations.
1358 */
1359 orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
1360 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1361 pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
1362 if (is_huge_zero_pmd(orig_pmd)) {
1363 atomic_long_dec(&tlb->mm->nr_ptes);
1364 spin_unlock(ptl);
1365 put_huge_zero_page();
1366 } else {
1367 page = pmd_page(orig_pmd);
1368 page_remove_rmap(page);
1369 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1370 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1371 VM_BUG_ON_PAGE(!PageHead(page), page);
1372 atomic_long_dec(&tlb->mm->nr_ptes);
1373 spin_unlock(ptl);
1374 tlb_remove_page(tlb, page);
1375 }
1376 pte_free(tlb->mm, pgtable);
1377 ret = 1;
1378 }
1379 return ret;
1380}
1381
1382int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1383 unsigned long addr, unsigned long end,
1384 unsigned char *vec)
1385{
1386 spinlock_t *ptl;
1387 int ret = 0;
1388
1389 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1390 /*
1391 * All logical pages in the range are present
1392 * if backed by a huge page.
1393 */
1394 spin_unlock(ptl);
1395 memset(vec, 1, (end - addr) >> PAGE_SHIFT);
1396 ret = 1;
1397 }
1398
1399 return ret;
1400}
1401
1402int move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1403 unsigned long old_addr,
1404 unsigned long new_addr, unsigned long old_end,
1405 pmd_t *old_pmd, pmd_t *new_pmd)
1406{
1407 spinlock_t *old_ptl, *new_ptl;
1408 int ret = 0;
1409 pmd_t pmd;
1410
1411 struct mm_struct *mm = vma->vm_mm;
1412
1413 if ((old_addr & ~HPAGE_PMD_MASK) ||
1414 (new_addr & ~HPAGE_PMD_MASK) ||
1415 old_end - old_addr < HPAGE_PMD_SIZE ||
1416 (new_vma->vm_flags & VM_NOHUGEPAGE))
1417 goto out;
1418
1419 /*
1420 * The destination pmd shouldn't be established, free_pgtables()
1421 * should have release it.
1422 */
1423 if (WARN_ON(!pmd_none(*new_pmd))) {
1424 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1425 goto out;
1426 }
1427
1428 /*
1429 * We don't have to worry about the ordering of src and dst
1430 * ptlocks because exclusive mmap_sem prevents deadlock.
1431 */
1432 ret = __pmd_trans_huge_lock(old_pmd, vma, &old_ptl);
1433 if (ret == 1) {
1434 new_ptl = pmd_lockptr(mm, new_pmd);
1435 if (new_ptl != old_ptl)
1436 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1437 pmd = pmdp_get_and_clear(mm, old_addr, old_pmd);
1438 VM_BUG_ON(!pmd_none(*new_pmd));
1439
1440 if (pmd_move_must_withdraw(new_ptl, old_ptl)) {
1441 pgtable_t pgtable;
1442 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1443 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1444 }
1445 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1446 if (new_ptl != old_ptl)
1447 spin_unlock(new_ptl);
1448 spin_unlock(old_ptl);
1449 }
1450out:
1451 return ret;
1452}
1453
1454/*
1455 * Returns
1456 * - 0 if PMD could not be locked
1457 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1458 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1459 */
1460int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1461 unsigned long addr, pgprot_t newprot, int prot_numa)
1462{
1463 struct mm_struct *mm = vma->vm_mm;
1464 spinlock_t *ptl;
1465 int ret = 0;
1466
1467 if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
1468 pmd_t entry;
1469 ret = 1;
1470 if (!prot_numa) {
1471 entry = pmdp_get_and_clear(mm, addr, pmd);
1472 if (pmd_numa(entry))
1473 entry = pmd_mknonnuma(entry);
1474 entry = pmd_modify(entry, newprot);
1475 ret = HPAGE_PMD_NR;
1476 set_pmd_at(mm, addr, pmd, entry);
1477 BUG_ON(pmd_write(entry));
1478 } else {
1479 struct page *page = pmd_page(*pmd);
1480
1481 /*
1482 * Do not trap faults against the zero page. The
1483 * read-only data is likely to be read-cached on the
1484 * local CPU cache and it is less useful to know about
1485 * local vs remote hits on the zero page.
1486 */
1487 if (!is_huge_zero_page(page) &&
1488 !pmd_numa(*pmd)) {
1489 pmdp_set_numa(mm, addr, pmd);
1490 ret = HPAGE_PMD_NR;
1491 }
1492 }
1493 spin_unlock(ptl);
1494 }
1495
1496 return ret;
1497}
1498
1499/*
1500 * Returns 1 if a given pmd maps a stable (not under splitting) thp.
1501 * Returns -1 if it maps a thp under splitting. Returns 0 otherwise.
1502 *
1503 * Note that if it returns 1, this routine returns without unlocking page
1504 * table locks. So callers must unlock them.
1505 */
1506int __pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma,
1507 spinlock_t **ptl)
1508{
1509 *ptl = pmd_lock(vma->vm_mm, pmd);
1510 if (likely(pmd_trans_huge(*pmd))) {
1511 if (unlikely(pmd_trans_splitting(*pmd))) {
1512 spin_unlock(*ptl);
1513 wait_split_huge_page(vma->anon_vma, pmd);
1514 return -1;
1515 } else {
1516 /* Thp mapped by 'pmd' is stable, so we can
1517 * handle it as it is. */
1518 return 1;
1519 }
1520 }
1521 spin_unlock(*ptl);
1522 return 0;
1523}
1524
1525/*
1526 * This function returns whether a given @page is mapped onto the @address
1527 * in the virtual space of @mm.
1528 *
1529 * When it's true, this function returns *pmd with holding the page table lock
1530 * and passing it back to the caller via @ptl.
1531 * If it's false, returns NULL without holding the page table lock.
1532 */
1533pmd_t *page_check_address_pmd(struct page *page,
1534 struct mm_struct *mm,
1535 unsigned long address,
1536 enum page_check_address_pmd_flag flag,
1537 spinlock_t **ptl)
1538{
1539 pgd_t *pgd;
1540 pud_t *pud;
1541 pmd_t *pmd;
1542
1543 if (address & ~HPAGE_PMD_MASK)
1544 return NULL;
1545
1546 pgd = pgd_offset(mm, address);
1547 if (!pgd_present(*pgd))
1548 return NULL;
1549 pud = pud_offset(pgd, address);
1550 if (!pud_present(*pud))
1551 return NULL;
1552 pmd = pmd_offset(pud, address);
1553
1554 *ptl = pmd_lock(mm, pmd);
1555 if (!pmd_present(*pmd))
1556 goto unlock;
1557 if (pmd_page(*pmd) != page)
1558 goto unlock;
1559 /*
1560 * split_vma() may create temporary aliased mappings. There is
1561 * no risk as long as all huge pmd are found and have their
1562 * splitting bit set before __split_huge_page_refcount
1563 * runs. Finding the same huge pmd more than once during the
1564 * same rmap walk is not a problem.
1565 */
1566 if (flag == PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG &&
1567 pmd_trans_splitting(*pmd))
1568 goto unlock;
1569 if (pmd_trans_huge(*pmd)) {
1570 VM_BUG_ON(flag == PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG &&
1571 !pmd_trans_splitting(*pmd));
1572 return pmd;
1573 }
1574unlock:
1575 spin_unlock(*ptl);
1576 return NULL;
1577}
1578
1579static int __split_huge_page_splitting(struct page *page,
1580 struct vm_area_struct *vma,
1581 unsigned long address)
1582{
1583 struct mm_struct *mm = vma->vm_mm;
1584 spinlock_t *ptl;
1585 pmd_t *pmd;
1586 int ret = 0;
1587 /* For mmu_notifiers */
1588 const unsigned long mmun_start = address;
1589 const unsigned long mmun_end = address + HPAGE_PMD_SIZE;
1590
1591 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1592 pmd = page_check_address_pmd(page, mm, address,
1593 PAGE_CHECK_ADDRESS_PMD_NOTSPLITTING_FLAG, &ptl);
1594 if (pmd) {
1595 /*
1596 * We can't temporarily set the pmd to null in order
1597 * to split it, the pmd must remain marked huge at all
1598 * times or the VM won't take the pmd_trans_huge paths
1599 * and it won't wait on the anon_vma->root->rwsem to
1600 * serialize against split_huge_page*.
1601 */
1602 pmdp_splitting_flush(vma, address, pmd);
1603 ret = 1;
1604 spin_unlock(ptl);
1605 }
1606 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1607
1608 return ret;
1609}
1610
1611static void __split_huge_page_refcount(struct page *page,
1612 struct list_head *list)
1613{
1614 int i;
1615 struct zone *zone = page_zone(page);
1616 struct lruvec *lruvec;
1617 int tail_count = 0;
1618
1619 /* prevent PageLRU to go away from under us, and freeze lru stats */
1620 spin_lock_irq(&zone->lru_lock);
1621 lruvec = mem_cgroup_page_lruvec(page, zone);
1622
1623 compound_lock(page);
1624 /* complete memcg works before add pages to LRU */
1625 mem_cgroup_split_huge_fixup(page);
1626
1627 for (i = HPAGE_PMD_NR - 1; i >= 1; i--) {
1628 struct page *page_tail = page + i;
1629
1630 /* tail_page->_mapcount cannot change */
1631 BUG_ON(page_mapcount(page_tail) < 0);
1632 tail_count += page_mapcount(page_tail);
1633 /* check for overflow */
1634 BUG_ON(tail_count < 0);
1635 BUG_ON(atomic_read(&page_tail->_count) != 0);
1636 /*
1637 * tail_page->_count is zero and not changing from
1638 * under us. But get_page_unless_zero() may be running
1639 * from under us on the tail_page. If we used
1640 * atomic_set() below instead of atomic_add(), we
1641 * would then run atomic_set() concurrently with
1642 * get_page_unless_zero(), and atomic_set() is
1643 * implemented in C not using locked ops. spin_unlock
1644 * on x86 sometime uses locked ops because of PPro
1645 * errata 66, 92, so unless somebody can guarantee
1646 * atomic_set() here would be safe on all archs (and
1647 * not only on x86), it's safer to use atomic_add().
1648 */
1649 atomic_add(page_mapcount(page) + page_mapcount(page_tail) + 1,
1650 &page_tail->_count);
1651
1652 /* after clearing PageTail the gup refcount can be released */
1653 smp_mb();
1654
1655 /*
1656 * retain hwpoison flag of the poisoned tail page:
1657 * fix for the unsuitable process killed on Guest Machine(KVM)
1658 * by the memory-failure.
1659 */
1660 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP | __PG_HWPOISON;
1661 page_tail->flags |= (page->flags &
1662 ((1L << PG_referenced) |
1663 (1L << PG_swapbacked) |
1664 (1L << PG_mlocked) |
1665 (1L << PG_uptodate) |
1666 (1L << PG_active) |
1667 (1L << PG_unevictable)));
1668 page_tail->flags |= (1L << PG_dirty);
1669
1670 /* clear PageTail before overwriting first_page */
1671 smp_wmb();
1672
1673 /*
1674 * __split_huge_page_splitting() already set the
1675 * splitting bit in all pmd that could map this
1676 * hugepage, that will ensure no CPU can alter the
1677 * mapcount on the head page. The mapcount is only
1678 * accounted in the head page and it has to be
1679 * transferred to all tail pages in the below code. So
1680 * for this code to be safe, the split the mapcount
1681 * can't change. But that doesn't mean userland can't
1682 * keep changing and reading the page contents while
1683 * we transfer the mapcount, so the pmd splitting
1684 * status is achieved setting a reserved bit in the
1685 * pmd, not by clearing the present bit.
1686 */
1687 page_tail->_mapcount = page->_mapcount;
1688
1689 BUG_ON(page_tail->mapping);
1690 page_tail->mapping = page->mapping;
1691
1692 page_tail->index = page->index + i;
1693 page_cpupid_xchg_last(page_tail, page_cpupid_last(page));
1694
1695 BUG_ON(!PageAnon(page_tail));
1696 BUG_ON(!PageUptodate(page_tail));
1697 BUG_ON(!PageDirty(page_tail));
1698 BUG_ON(!PageSwapBacked(page_tail));
1699
1700 lru_add_page_tail(page, page_tail, lruvec, list);
1701 }
1702 atomic_sub(tail_count, &page->_count);
1703 BUG_ON(atomic_read(&page->_count) <= 0);
1704
1705 __mod_zone_page_state(zone, NR_ANON_TRANSPARENT_HUGEPAGES, -1);
1706
1707 ClearPageCompound(page);
1708 compound_unlock(page);
1709 spin_unlock_irq(&zone->lru_lock);
1710
1711 for (i = 1; i < HPAGE_PMD_NR; i++) {
1712 struct page *page_tail = page + i;
1713 BUG_ON(page_count(page_tail) <= 0);
1714 /*
1715 * Tail pages may be freed if there wasn't any mapping
1716 * like if add_to_swap() is running on a lru page that
1717 * had its mapping zapped. And freeing these pages
1718 * requires taking the lru_lock so we do the put_page
1719 * of the tail pages after the split is complete.
1720 */
1721 put_page(page_tail);
1722 }
1723
1724 /*
1725 * Only the head page (now become a regular page) is required
1726 * to be pinned by the caller.
1727 */
1728 BUG_ON(page_count(page) <= 0);
1729}
1730
1731static int __split_huge_page_map(struct page *page,
1732 struct vm_area_struct *vma,
1733 unsigned long address)
1734{
1735 struct mm_struct *mm = vma->vm_mm;
1736 spinlock_t *ptl;
1737 pmd_t *pmd, _pmd;
1738 int ret = 0, i;
1739 pgtable_t pgtable;
1740 unsigned long haddr;
1741
1742 pmd = page_check_address_pmd(page, mm, address,
1743 PAGE_CHECK_ADDRESS_PMD_SPLITTING_FLAG, &ptl);
1744 if (pmd) {
1745 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1746 pmd_populate(mm, &_pmd, pgtable);
1747
1748 haddr = address;
1749 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1750 pte_t *pte, entry;
1751 BUG_ON(PageCompound(page+i));
1752 entry = mk_pte(page + i, vma->vm_page_prot);
1753 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1754 if (!pmd_write(*pmd))
1755 entry = pte_wrprotect(entry);
1756 else
1757 BUG_ON(page_mapcount(page) != 1);
1758 if (!pmd_young(*pmd))
1759 entry = pte_mkold(entry);
1760 if (pmd_numa(*pmd))
1761 entry = pte_mknuma(entry);
1762 pte = pte_offset_map(&_pmd, haddr);
1763 BUG_ON(!pte_none(*pte));
1764 set_pte_at(mm, haddr, pte, entry);
1765 pte_unmap(pte);
1766 }
1767
1768 smp_wmb(); /* make pte visible before pmd */
1769 /*
1770 * Up to this point the pmd is present and huge and
1771 * userland has the whole access to the hugepage
1772 * during the split (which happens in place). If we
1773 * overwrite the pmd with the not-huge version
1774 * pointing to the pte here (which of course we could
1775 * if all CPUs were bug free), userland could trigger
1776 * a small page size TLB miss on the small sized TLB
1777 * while the hugepage TLB entry is still established
1778 * in the huge TLB. Some CPU doesn't like that. See
1779 * http://support.amd.com/us/Processor_TechDocs/41322.pdf,
1780 * Erratum 383 on page 93. Intel should be safe but is
1781 * also warns that it's only safe if the permission
1782 * and cache attributes of the two entries loaded in
1783 * the two TLB is identical (which should be the case
1784 * here). But it is generally safer to never allow
1785 * small and huge TLB entries for the same virtual
1786 * address to be loaded simultaneously. So instead of
1787 * doing "pmd_populate(); flush_tlb_range();" we first
1788 * mark the current pmd notpresent (atomically because
1789 * here the pmd_trans_huge and pmd_trans_splitting
1790 * must remain set at all times on the pmd until the
1791 * split is complete for this pmd), then we flush the
1792 * SMP TLB and finally we write the non-huge version
1793 * of the pmd entry with pmd_populate.
1794 */
1795 pmdp_invalidate(vma, address, pmd);
1796 pmd_populate(mm, pmd, pgtable);
1797 ret = 1;
1798 spin_unlock(ptl);
1799 }
1800
1801 return ret;
1802}
1803
1804/* must be called with anon_vma->root->rwsem held */
1805static void __split_huge_page(struct page *page,
1806 struct anon_vma *anon_vma,
1807 struct list_head *list)
1808{
1809 int mapcount, mapcount2;
1810 pgoff_t pgoff = page->index << (PAGE_CACHE_SHIFT - PAGE_SHIFT);
1811 struct anon_vma_chain *avc;
1812
1813 BUG_ON(!PageHead(page));
1814 BUG_ON(PageTail(page));
1815
1816 mapcount = 0;
1817 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1818 struct vm_area_struct *vma = avc->vma;
1819 unsigned long addr = vma_address(page, vma);
1820 BUG_ON(is_vma_temporary_stack(vma));
1821 mapcount += __split_huge_page_splitting(page, vma, addr);
1822 }
1823 /*
1824 * It is critical that new vmas are added to the tail of the
1825 * anon_vma list. This guarantes that if copy_huge_pmd() runs
1826 * and establishes a child pmd before
1827 * __split_huge_page_splitting() freezes the parent pmd (so if
1828 * we fail to prevent copy_huge_pmd() from running until the
1829 * whole __split_huge_page() is complete), we will still see
1830 * the newly established pmd of the child later during the
1831 * walk, to be able to set it as pmd_trans_splitting too.
1832 */
1833 if (mapcount != page_mapcount(page))
1834 printk(KERN_ERR "mapcount %d page_mapcount %d\n",
1835 mapcount, page_mapcount(page));
1836 BUG_ON(mapcount != page_mapcount(page));
1837
1838 __split_huge_page_refcount(page, list);
1839
1840 mapcount2 = 0;
1841 anon_vma_interval_tree_foreach(avc, &anon_vma->rb_root, pgoff, pgoff) {
1842 struct vm_area_struct *vma = avc->vma;
1843 unsigned long addr = vma_address(page, vma);
1844 BUG_ON(is_vma_temporary_stack(vma));
1845 mapcount2 += __split_huge_page_map(page, vma, addr);
1846 }
1847 if (mapcount != mapcount2)
1848 printk(KERN_ERR "mapcount %d mapcount2 %d page_mapcount %d\n",
1849 mapcount, mapcount2, page_mapcount(page));
1850 BUG_ON(mapcount != mapcount2);
1851}
1852
1853/*
1854 * Split a hugepage into normal pages. This doesn't change the position of head
1855 * page. If @list is null, tail pages will be added to LRU list, otherwise, to
1856 * @list. Both head page and tail pages will inherit mapping, flags, and so on
1857 * from the hugepage.
1858 * Return 0 if the hugepage is split successfully otherwise return 1.
1859 */
1860int split_huge_page_to_list(struct page *page, struct list_head *list)
1861{
1862 struct anon_vma *anon_vma;
1863 int ret = 1;
1864
1865 BUG_ON(is_huge_zero_page(page));
1866 BUG_ON(!PageAnon(page));
1867
1868 /*
1869 * The caller does not necessarily hold an mmap_sem that would prevent
1870 * the anon_vma disappearing so we first we take a reference to it
1871 * and then lock the anon_vma for write. This is similar to
1872 * page_lock_anon_vma_read except the write lock is taken to serialise
1873 * against parallel split or collapse operations.
1874 */
1875 anon_vma = page_get_anon_vma(page);
1876 if (!anon_vma)
1877 goto out;
1878 anon_vma_lock_write(anon_vma);
1879
1880 ret = 0;
1881 if (!PageCompound(page))
1882 goto out_unlock;
1883
1884 BUG_ON(!PageSwapBacked(page));
1885 __split_huge_page(page, anon_vma, list);
1886 count_vm_event(THP_SPLIT);
1887
1888 BUG_ON(PageCompound(page));
1889out_unlock:
1890 anon_vma_unlock_write(anon_vma);
1891 put_anon_vma(anon_vma);
1892out:
1893 return ret;
1894}
1895
1896#define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
1897
1898int hugepage_madvise(struct vm_area_struct *vma,
1899 unsigned long *vm_flags, int advice)
1900{
1901 switch (advice) {
1902 case MADV_HUGEPAGE:
1903#ifdef CONFIG_S390
1904 /*
1905 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1906 * can't handle this properly after s390_enable_sie, so we simply
1907 * ignore the madvise to prevent qemu from causing a SIGSEGV.
1908 */
1909 if (mm_has_pgste(vma->vm_mm))
1910 return 0;
1911#endif
1912 /*
1913 * Be somewhat over-protective like KSM for now!
1914 */
1915 if (*vm_flags & (VM_HUGEPAGE | VM_NO_THP))
1916 return -EINVAL;
1917 *vm_flags &= ~VM_NOHUGEPAGE;
1918 *vm_flags |= VM_HUGEPAGE;
1919 /*
1920 * If the vma become good for khugepaged to scan,
1921 * register it here without waiting a page fault that
1922 * may not happen any time soon.
1923 */
1924 if (unlikely(khugepaged_enter_vma_merge(vma)))
1925 return -ENOMEM;
1926 break;
1927 case MADV_NOHUGEPAGE:
1928 /*
1929 * Be somewhat over-protective like KSM for now!
1930 */
1931 if (*vm_flags & (VM_NOHUGEPAGE | VM_NO_THP))
1932 return -EINVAL;
1933 *vm_flags &= ~VM_HUGEPAGE;
1934 *vm_flags |= VM_NOHUGEPAGE;
1935 /*
1936 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1937 * this vma even if we leave the mm registered in khugepaged if
1938 * it got registered before VM_NOHUGEPAGE was set.
1939 */
1940 break;
1941 }
1942
1943 return 0;
1944}
1945
1946static int __init khugepaged_slab_init(void)
1947{
1948 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1949 sizeof(struct mm_slot),
1950 __alignof__(struct mm_slot), 0, NULL);
1951 if (!mm_slot_cache)
1952 return -ENOMEM;
1953
1954 return 0;
1955}
1956
1957static inline struct mm_slot *alloc_mm_slot(void)
1958{
1959 if (!mm_slot_cache) /* initialization failed */
1960 return NULL;
1961 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1962}
1963
1964static inline void free_mm_slot(struct mm_slot *mm_slot)
1965{
1966 kmem_cache_free(mm_slot_cache, mm_slot);
1967}
1968
1969static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1970{
1971 struct mm_slot *mm_slot;
1972
1973 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
1974 if (mm == mm_slot->mm)
1975 return mm_slot;
1976
1977 return NULL;
1978}
1979
1980static void insert_to_mm_slots_hash(struct mm_struct *mm,
1981 struct mm_slot *mm_slot)
1982{
1983 mm_slot->mm = mm;
1984 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
1985}
1986
1987static inline int khugepaged_test_exit(struct mm_struct *mm)
1988{
1989 return atomic_read(&mm->mm_users) == 0;
1990}
1991
1992int __khugepaged_enter(struct mm_struct *mm)
1993{
1994 struct mm_slot *mm_slot;
1995 int wakeup;
1996
1997 mm_slot = alloc_mm_slot();
1998 if (!mm_slot)
1999 return -ENOMEM;
2000
2001 /* __khugepaged_exit() must not run from under us */
2002 VM_BUG_ON(khugepaged_test_exit(mm));
2003 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
2004 free_mm_slot(mm_slot);
2005 return 0;
2006 }
2007
2008 spin_lock(&khugepaged_mm_lock);
2009 insert_to_mm_slots_hash(mm, mm_slot);
2010 /*
2011 * Insert just behind the scanning cursor, to let the area settle
2012 * down a little.
2013 */
2014 wakeup = list_empty(&khugepaged_scan.mm_head);
2015 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
2016 spin_unlock(&khugepaged_mm_lock);
2017
2018 atomic_inc(&mm->mm_count);
2019 if (wakeup)
2020 wake_up_interruptible(&khugepaged_wait);
2021
2022 return 0;
2023}
2024
2025int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
2026{
2027 unsigned long hstart, hend;
2028 if (!vma->anon_vma)
2029 /*
2030 * Not yet faulted in so we will register later in the
2031 * page fault if needed.
2032 */
2033 return 0;
2034 if (vma->vm_ops)
2035 /* khugepaged not yet working on file or special mappings */
2036 return 0;
2037 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2038 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2039 hend = vma->vm_end & HPAGE_PMD_MASK;
2040 if (hstart < hend)
2041 return khugepaged_enter(vma);
2042 return 0;
2043}
2044
2045void __khugepaged_exit(struct mm_struct *mm)
2046{
2047 struct mm_slot *mm_slot;
2048 int free = 0;
2049
2050 spin_lock(&khugepaged_mm_lock);
2051 mm_slot = get_mm_slot(mm);
2052 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
2053 hash_del(&mm_slot->hash);
2054 list_del(&mm_slot->mm_node);
2055 free = 1;
2056 }
2057 spin_unlock(&khugepaged_mm_lock);
2058
2059 if (free) {
2060 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2061 free_mm_slot(mm_slot);
2062 mmdrop(mm);
2063 } else if (mm_slot) {
2064 /*
2065 * This is required to serialize against
2066 * khugepaged_test_exit() (which is guaranteed to run
2067 * under mmap sem read mode). Stop here (after we
2068 * return all pagetables will be destroyed) until
2069 * khugepaged has finished working on the pagetables
2070 * under the mmap_sem.
2071 */
2072 down_write(&mm->mmap_sem);
2073 up_write(&mm->mmap_sem);
2074 }
2075}
2076
2077static void release_pte_page(struct page *page)
2078{
2079 /* 0 stands for page_is_file_cache(page) == false */
2080 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2081 unlock_page(page);
2082 putback_lru_page(page);
2083}
2084
2085static void release_pte_pages(pte_t *pte, pte_t *_pte)
2086{
2087 while (--_pte >= pte) {
2088 pte_t pteval = *_pte;
2089 if (!pte_none(pteval))
2090 release_pte_page(pte_page(pteval));
2091 }
2092}
2093
2094static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2095 unsigned long address,
2096 pte_t *pte)
2097{
2098 struct page *page;
2099 pte_t *_pte;
2100 int referenced = 0, none = 0;
2101 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2102 _pte++, address += PAGE_SIZE) {
2103 pte_t pteval = *_pte;
2104 if (pte_none(pteval)) {
2105 if (++none <= khugepaged_max_ptes_none)
2106 continue;
2107 else
2108 goto out;
2109 }
2110 if (!pte_present(pteval) || !pte_write(pteval))
2111 goto out;
2112 page = vm_normal_page(vma, address, pteval);
2113 if (unlikely(!page))
2114 goto out;
2115
2116 VM_BUG_ON_PAGE(PageCompound(page), page);
2117 VM_BUG_ON_PAGE(!PageAnon(page), page);
2118 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2119
2120 /* cannot use mapcount: can't collapse if there's a gup pin */
2121 if (page_count(page) != 1)
2122 goto out;
2123 /*
2124 * We can do it before isolate_lru_page because the
2125 * page can't be freed from under us. NOTE: PG_lock
2126 * is needed to serialize against split_huge_page
2127 * when invoked from the VM.
2128 */
2129 if (!trylock_page(page))
2130 goto out;
2131 /*
2132 * Isolate the page to avoid collapsing an hugepage
2133 * currently in use by the VM.
2134 */
2135 if (isolate_lru_page(page)) {
2136 unlock_page(page);
2137 goto out;
2138 }
2139 /* 0 stands for page_is_file_cache(page) == false */
2140 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2141 VM_BUG_ON_PAGE(!PageLocked(page), page);
2142 VM_BUG_ON_PAGE(PageLRU(page), page);
2143
2144 /* If there is no mapped pte young don't collapse the page */
2145 if (pte_young(pteval) || PageReferenced(page) ||
2146 mmu_notifier_test_young(vma->vm_mm, address))
2147 referenced = 1;
2148 }
2149 if (likely(referenced))
2150 return 1;
2151out:
2152 release_pte_pages(pte, _pte);
2153 return 0;
2154}
2155
2156static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2157 struct vm_area_struct *vma,
2158 unsigned long address,
2159 spinlock_t *ptl)
2160{
2161 pte_t *_pte;
2162 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2163 pte_t pteval = *_pte;
2164 struct page *src_page;
2165
2166 if (pte_none(pteval)) {
2167 clear_user_highpage(page, address);
2168 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2169 } else {
2170 src_page = pte_page(pteval);
2171 copy_user_highpage(page, src_page, address, vma);
2172 VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2173 release_pte_page(src_page);
2174 /*
2175 * ptl mostly unnecessary, but preempt has to
2176 * be disabled to update the per-cpu stats
2177 * inside page_remove_rmap().
2178 */
2179 spin_lock(ptl);
2180 /*
2181 * paravirt calls inside pte_clear here are
2182 * superfluous.
2183 */
2184 pte_clear(vma->vm_mm, address, _pte);
2185 page_remove_rmap(src_page);
2186 spin_unlock(ptl);
2187 free_page_and_swap_cache(src_page);
2188 }
2189
2190 address += PAGE_SIZE;
2191 page++;
2192 }
2193}
2194
2195static void khugepaged_alloc_sleep(void)
2196{
2197 wait_event_freezable_timeout(khugepaged_wait, false,
2198 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2199}
2200
2201static int khugepaged_node_load[MAX_NUMNODES];
2202
2203#ifdef CONFIG_NUMA
2204static int khugepaged_find_target_node(void)
2205{
2206 static int last_khugepaged_target_node = NUMA_NO_NODE;
2207 int nid, target_node = 0, max_value = 0;
2208
2209 /* find first node with max normal pages hit */
2210 for (nid = 0; nid < MAX_NUMNODES; nid++)
2211 if (khugepaged_node_load[nid] > max_value) {
2212 max_value = khugepaged_node_load[nid];
2213 target_node = nid;
2214 }
2215
2216 /* do some balance if several nodes have the same hit record */
2217 if (target_node <= last_khugepaged_target_node)
2218 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2219 nid++)
2220 if (max_value == khugepaged_node_load[nid]) {
2221 target_node = nid;
2222 break;
2223 }
2224
2225 last_khugepaged_target_node = target_node;
2226 return target_node;
2227}
2228
2229static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2230{
2231 if (IS_ERR(*hpage)) {
2232 if (!*wait)
2233 return false;
2234
2235 *wait = false;
2236 *hpage = NULL;
2237 khugepaged_alloc_sleep();
2238 } else if (*hpage) {
2239 put_page(*hpage);
2240 *hpage = NULL;
2241 }
2242
2243 return true;
2244}
2245
2246static struct page
2247*khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2248 struct vm_area_struct *vma, unsigned long address,
2249 int node)
2250{
2251 VM_BUG_ON_PAGE(*hpage, *hpage);
2252 /*
2253 * Allocate the page while the vma is still valid and under
2254 * the mmap_sem read mode so there is no memory allocation
2255 * later when we take the mmap_sem in write mode. This is more
2256 * friendly behavior (OTOH it may actually hide bugs) to
2257 * filesystems in userland with daemons allocating memory in
2258 * the userland I/O paths. Allocating memory with the
2259 * mmap_sem in read mode is good idea also to allow greater
2260 * scalability.
2261 */
2262 *hpage = alloc_pages_exact_node(node, alloc_hugepage_gfpmask(
2263 khugepaged_defrag(), __GFP_OTHER_NODE), HPAGE_PMD_ORDER);
2264 /*
2265 * After allocating the hugepage, release the mmap_sem read lock in
2266 * preparation for taking it in write mode.
2267 */
2268 up_read(&mm->mmap_sem);
2269 if (unlikely(!*hpage)) {
2270 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2271 *hpage = ERR_PTR(-ENOMEM);
2272 return NULL;
2273 }
2274
2275 count_vm_event(THP_COLLAPSE_ALLOC);
2276 return *hpage;
2277}
2278#else
2279static int khugepaged_find_target_node(void)
2280{
2281 return 0;
2282}
2283
2284static inline struct page *alloc_hugepage(int defrag)
2285{
2286 return alloc_pages(alloc_hugepage_gfpmask(defrag, 0),
2287 HPAGE_PMD_ORDER);
2288}
2289
2290static struct page *khugepaged_alloc_hugepage(bool *wait)
2291{
2292 struct page *hpage;
2293
2294 do {
2295 hpage = alloc_hugepage(khugepaged_defrag());
2296 if (!hpage) {
2297 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2298 if (!*wait)
2299 return NULL;
2300
2301 *wait = false;
2302 khugepaged_alloc_sleep();
2303 } else
2304 count_vm_event(THP_COLLAPSE_ALLOC);
2305 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2306
2307 return hpage;
2308}
2309
2310static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2311{
2312 if (!*hpage)
2313 *hpage = khugepaged_alloc_hugepage(wait);
2314
2315 if (unlikely(!*hpage))
2316 return false;
2317
2318 return true;
2319}
2320
2321static struct page
2322*khugepaged_alloc_page(struct page **hpage, struct mm_struct *mm,
2323 struct vm_area_struct *vma, unsigned long address,
2324 int node)
2325{
2326 up_read(&mm->mmap_sem);
2327 VM_BUG_ON(!*hpage);
2328 return *hpage;
2329}
2330#endif
2331
2332static bool hugepage_vma_check(struct vm_area_struct *vma)
2333{
2334 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2335 (vma->vm_flags & VM_NOHUGEPAGE))
2336 return false;
2337
2338 if (!vma->anon_vma || vma->vm_ops)
2339 return false;
2340 if (is_vma_temporary_stack(vma))
2341 return false;
2342 VM_BUG_ON(vma->vm_flags & VM_NO_THP);
2343 return true;
2344}
2345
2346static void collapse_huge_page(struct mm_struct *mm,
2347 unsigned long address,
2348 struct page **hpage,
2349 struct vm_area_struct *vma,
2350 int node)
2351{
2352 pmd_t *pmd, _pmd;
2353 pte_t *pte;
2354 pgtable_t pgtable;
2355 struct page *new_page;
2356 spinlock_t *pmd_ptl, *pte_ptl;
2357 int isolated;
2358 unsigned long hstart, hend;
2359 unsigned long mmun_start; /* For mmu_notifiers */
2360 unsigned long mmun_end; /* For mmu_notifiers */
2361
2362 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2363
2364 /* release the mmap_sem read lock. */
2365 new_page = khugepaged_alloc_page(hpage, mm, vma, address, node);
2366 if (!new_page)
2367 return;
2368
2369 if (unlikely(mem_cgroup_charge_anon(new_page, mm, GFP_KERNEL)))
2370 return;
2371
2372 /*
2373 * Prevent all access to pagetables with the exception of
2374 * gup_fast later hanlded by the ptep_clear_flush and the VM
2375 * handled by the anon_vma lock + PG_lock.
2376 */
2377 down_write(&mm->mmap_sem);
2378 if (unlikely(khugepaged_test_exit(mm)))
2379 goto out;
2380
2381 vma = find_vma(mm, address);
2382 if (!vma)
2383 goto out;
2384 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2385 hend = vma->vm_end & HPAGE_PMD_MASK;
2386 if (address < hstart || address + HPAGE_PMD_SIZE > hend)
2387 goto out;
2388 if (!hugepage_vma_check(vma))
2389 goto out;
2390 pmd = mm_find_pmd(mm, address);
2391 if (!pmd)
2392 goto out;
2393 if (pmd_trans_huge(*pmd))
2394 goto out;
2395
2396 anon_vma_lock_write(vma->anon_vma);
2397
2398 pte = pte_offset_map(pmd, address);
2399 pte_ptl = pte_lockptr(mm, pmd);
2400
2401 mmun_start = address;
2402 mmun_end = address + HPAGE_PMD_SIZE;
2403 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2404 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2405 /*
2406 * After this gup_fast can't run anymore. This also removes
2407 * any huge TLB entry from the CPU so we won't allow
2408 * huge and small TLB entries for the same virtual address
2409 * to avoid the risk of CPU bugs in that area.
2410 */
2411 _pmd = pmdp_clear_flush(vma, address, pmd);
2412 spin_unlock(pmd_ptl);
2413 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2414
2415 spin_lock(pte_ptl);
2416 isolated = __collapse_huge_page_isolate(vma, address, pte);
2417 spin_unlock(pte_ptl);
2418
2419 if (unlikely(!isolated)) {
2420 pte_unmap(pte);
2421 spin_lock(pmd_ptl);
2422 BUG_ON(!pmd_none(*pmd));
2423 /*
2424 * We can only use set_pmd_at when establishing
2425 * hugepmds and never for establishing regular pmds that
2426 * points to regular pagetables. Use pmd_populate for that
2427 */
2428 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2429 spin_unlock(pmd_ptl);
2430 anon_vma_unlock_write(vma->anon_vma);
2431 goto out;
2432 }
2433
2434 /*
2435 * All pages are isolated and locked so anon_vma rmap
2436 * can't run anymore.
2437 */
2438 anon_vma_unlock_write(vma->anon_vma);
2439
2440 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2441 pte_unmap(pte);
2442 __SetPageUptodate(new_page);
2443 pgtable = pmd_pgtable(_pmd);
2444
2445 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2446 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2447
2448 /*
2449 * spin_lock() below is not the equivalent of smp_wmb(), so
2450 * this is needed to avoid the copy_huge_page writes to become
2451 * visible after the set_pmd_at() write.
2452 */
2453 smp_wmb();
2454
2455 spin_lock(pmd_ptl);
2456 BUG_ON(!pmd_none(*pmd));
2457 page_add_new_anon_rmap(new_page, vma, address);
2458 pgtable_trans_huge_deposit(mm, pmd, pgtable);
2459 set_pmd_at(mm, address, pmd, _pmd);
2460 update_mmu_cache_pmd(vma, address, pmd);
2461 spin_unlock(pmd_ptl);
2462
2463 *hpage = NULL;
2464
2465 khugepaged_pages_collapsed++;
2466out_up_write:
2467 up_write(&mm->mmap_sem);
2468 return;
2469
2470out:
2471 mem_cgroup_uncharge_page(new_page);
2472 goto out_up_write;
2473}
2474
2475static int khugepaged_scan_pmd(struct mm_struct *mm,
2476 struct vm_area_struct *vma,
2477 unsigned long address,
2478 struct page **hpage)
2479{
2480 pmd_t *pmd;
2481 pte_t *pte, *_pte;
2482 int ret = 0, referenced = 0, none = 0;
2483 struct page *page;
2484 unsigned long _address;
2485 spinlock_t *ptl;
2486 int node = NUMA_NO_NODE;
2487
2488 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2489
2490 pmd = mm_find_pmd(mm, address);
2491 if (!pmd)
2492 goto out;
2493 if (pmd_trans_huge(*pmd))
2494 goto out;
2495
2496 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2497 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2498 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2499 _pte++, _address += PAGE_SIZE) {
2500 pte_t pteval = *_pte;
2501 if (pte_none(pteval)) {
2502 if (++none <= khugepaged_max_ptes_none)
2503 continue;
2504 else
2505 goto out_unmap;
2506 }
2507 if (!pte_present(pteval) || !pte_write(pteval))
2508 goto out_unmap;
2509 page = vm_normal_page(vma, _address, pteval);
2510 if (unlikely(!page))
2511 goto out_unmap;
2512 /*
2513 * Record which node the original page is from and save this
2514 * information to khugepaged_node_load[].
2515 * Khupaged will allocate hugepage from the node has the max
2516 * hit record.
2517 */
2518 node = page_to_nid(page);
2519 khugepaged_node_load[node]++;
2520 VM_BUG_ON_PAGE(PageCompound(page), page);
2521 if (!PageLRU(page) || PageLocked(page) || !PageAnon(page))
2522 goto out_unmap;
2523 /* cannot use mapcount: can't collapse if there's a gup pin */
2524 if (page_count(page) != 1)
2525 goto out_unmap;
2526 if (pte_young(pteval) || PageReferenced(page) ||
2527 mmu_notifier_test_young(vma->vm_mm, address))
2528 referenced = 1;
2529 }
2530 if (referenced)
2531 ret = 1;
2532out_unmap:
2533 pte_unmap_unlock(pte, ptl);
2534 if (ret) {
2535 node = khugepaged_find_target_node();
2536 /* collapse_huge_page will return with the mmap_sem released */
2537 collapse_huge_page(mm, address, hpage, vma, node);
2538 }
2539out:
2540 return ret;
2541}
2542
2543static void collect_mm_slot(struct mm_slot *mm_slot)
2544{
2545 struct mm_struct *mm = mm_slot->mm;
2546
2547 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2548
2549 if (khugepaged_test_exit(mm)) {
2550 /* free mm_slot */
2551 hash_del(&mm_slot->hash);
2552 list_del(&mm_slot->mm_node);
2553
2554 /*
2555 * Not strictly needed because the mm exited already.
2556 *
2557 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2558 */
2559
2560 /* khugepaged_mm_lock actually not necessary for the below */
2561 free_mm_slot(mm_slot);
2562 mmdrop(mm);
2563 }
2564}
2565
2566static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2567 struct page **hpage)
2568 __releases(&khugepaged_mm_lock)
2569 __acquires(&khugepaged_mm_lock)
2570{
2571 struct mm_slot *mm_slot;
2572 struct mm_struct *mm;
2573 struct vm_area_struct *vma;
2574 int progress = 0;
2575
2576 VM_BUG_ON(!pages);
2577 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2578
2579 if (khugepaged_scan.mm_slot)
2580 mm_slot = khugepaged_scan.mm_slot;
2581 else {
2582 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2583 struct mm_slot, mm_node);
2584 khugepaged_scan.address = 0;
2585 khugepaged_scan.mm_slot = mm_slot;
2586 }
2587 spin_unlock(&khugepaged_mm_lock);
2588
2589 mm = mm_slot->mm;
2590 down_read(&mm->mmap_sem);
2591 if (unlikely(khugepaged_test_exit(mm)))
2592 vma = NULL;
2593 else
2594 vma = find_vma(mm, khugepaged_scan.address);
2595
2596 progress++;
2597 for (; vma; vma = vma->vm_next) {
2598 unsigned long hstart, hend;
2599
2600 cond_resched();
2601 if (unlikely(khugepaged_test_exit(mm))) {
2602 progress++;
2603 break;
2604 }
2605 if (!hugepage_vma_check(vma)) {
2606skip:
2607 progress++;
2608 continue;
2609 }
2610 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2611 hend = vma->vm_end & HPAGE_PMD_MASK;
2612 if (hstart >= hend)
2613 goto skip;
2614 if (khugepaged_scan.address > hend)
2615 goto skip;
2616 if (khugepaged_scan.address < hstart)
2617 khugepaged_scan.address = hstart;
2618 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2619
2620 while (khugepaged_scan.address < hend) {
2621 int ret;
2622 cond_resched();
2623 if (unlikely(khugepaged_test_exit(mm)))
2624 goto breakouterloop;
2625
2626 VM_BUG_ON(khugepaged_scan.address < hstart ||
2627 khugepaged_scan.address + HPAGE_PMD_SIZE >
2628 hend);
2629 ret = khugepaged_scan_pmd(mm, vma,
2630 khugepaged_scan.address,
2631 hpage);
2632 /* move to next address */
2633 khugepaged_scan.address += HPAGE_PMD_SIZE;
2634 progress += HPAGE_PMD_NR;
2635 if (ret)
2636 /* we released mmap_sem so break loop */
2637 goto breakouterloop_mmap_sem;
2638 if (progress >= pages)
2639 goto breakouterloop;
2640 }
2641 }
2642breakouterloop:
2643 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2644breakouterloop_mmap_sem:
2645
2646 spin_lock(&khugepaged_mm_lock);
2647 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2648 /*
2649 * Release the current mm_slot if this mm is about to die, or
2650 * if we scanned all vmas of this mm.
2651 */
2652 if (khugepaged_test_exit(mm) || !vma) {
2653 /*
2654 * Make sure that if mm_users is reaching zero while
2655 * khugepaged runs here, khugepaged_exit will find
2656 * mm_slot not pointing to the exiting mm.
2657 */
2658 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2659 khugepaged_scan.mm_slot = list_entry(
2660 mm_slot->mm_node.next,
2661 struct mm_slot, mm_node);
2662 khugepaged_scan.address = 0;
2663 } else {
2664 khugepaged_scan.mm_slot = NULL;
2665 khugepaged_full_scans++;
2666 }
2667
2668 collect_mm_slot(mm_slot);
2669 }
2670
2671 return progress;
2672}
2673
2674static int khugepaged_has_work(void)
2675{
2676 return !list_empty(&khugepaged_scan.mm_head) &&
2677 khugepaged_enabled();
2678}
2679
2680static int khugepaged_wait_event(void)
2681{
2682 return !list_empty(&khugepaged_scan.mm_head) ||
2683 kthread_should_stop();
2684}
2685
2686static void khugepaged_do_scan(void)
2687{
2688 struct page *hpage = NULL;
2689 unsigned int progress = 0, pass_through_head = 0;
2690 unsigned int pages = khugepaged_pages_to_scan;
2691 bool wait = true;
2692
2693 barrier(); /* write khugepaged_pages_to_scan to local stack */
2694
2695 while (progress < pages) {
2696 if (!khugepaged_prealloc_page(&hpage, &wait))
2697 break;
2698
2699 cond_resched();
2700
2701 if (unlikely(kthread_should_stop() || freezing(current)))
2702 break;
2703
2704 spin_lock(&khugepaged_mm_lock);
2705 if (!khugepaged_scan.mm_slot)
2706 pass_through_head++;
2707 if (khugepaged_has_work() &&
2708 pass_through_head < 2)
2709 progress += khugepaged_scan_mm_slot(pages - progress,
2710 &hpage);
2711 else
2712 progress = pages;
2713 spin_unlock(&khugepaged_mm_lock);
2714 }
2715
2716 if (!IS_ERR_OR_NULL(hpage))
2717 put_page(hpage);
2718}
2719
2720static void khugepaged_wait_work(void)
2721{
2722 try_to_freeze();
2723
2724 if (khugepaged_has_work()) {
2725 if (!khugepaged_scan_sleep_millisecs)
2726 return;
2727
2728 wait_event_freezable_timeout(khugepaged_wait,
2729 kthread_should_stop(),
2730 msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2731 return;
2732 }
2733
2734 if (khugepaged_enabled())
2735 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2736}
2737
2738static int khugepaged(void *none)
2739{
2740 struct mm_slot *mm_slot;
2741
2742 set_freezable();
2743 set_user_nice(current, 19);
2744
2745 while (!kthread_should_stop()) {
2746 khugepaged_do_scan();
2747 khugepaged_wait_work();
2748 }
2749
2750 spin_lock(&khugepaged_mm_lock);
2751 mm_slot = khugepaged_scan.mm_slot;
2752 khugepaged_scan.mm_slot = NULL;
2753 if (mm_slot)
2754 collect_mm_slot(mm_slot);
2755 spin_unlock(&khugepaged_mm_lock);
2756 return 0;
2757}
2758
2759static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2760 unsigned long haddr, pmd_t *pmd)
2761{
2762 struct mm_struct *mm = vma->vm_mm;
2763 pgtable_t pgtable;
2764 pmd_t _pmd;
2765 int i;
2766
2767 pmdp_clear_flush(vma, haddr, pmd);
2768 /* leave pmd empty until pte is filled */
2769
2770 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2771 pmd_populate(mm, &_pmd, pgtable);
2772
2773 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2774 pte_t *pte, entry;
2775 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2776 entry = pte_mkspecial(entry);
2777 pte = pte_offset_map(&_pmd, haddr);
2778 VM_BUG_ON(!pte_none(*pte));
2779 set_pte_at(mm, haddr, pte, entry);
2780 pte_unmap(pte);
2781 }
2782 smp_wmb(); /* make pte visible before pmd */
2783 pmd_populate(mm, pmd, pgtable);
2784 put_huge_zero_page();
2785}
2786
2787void __split_huge_page_pmd(struct vm_area_struct *vma, unsigned long address,
2788 pmd_t *pmd)
2789{
2790 spinlock_t *ptl;
2791 struct page *page;
2792 struct mm_struct *mm = vma->vm_mm;
2793 unsigned long haddr = address & HPAGE_PMD_MASK;
2794 unsigned long mmun_start; /* For mmu_notifiers */
2795 unsigned long mmun_end; /* For mmu_notifiers */
2796
2797 BUG_ON(vma->vm_start > haddr || vma->vm_end < haddr + HPAGE_PMD_SIZE);
2798
2799 mmun_start = haddr;
2800 mmun_end = haddr + HPAGE_PMD_SIZE;
2801again:
2802 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2803 ptl = pmd_lock(mm, pmd);
2804 if (unlikely(!pmd_trans_huge(*pmd))) {
2805 spin_unlock(ptl);
2806 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2807 return;
2808 }
2809 if (is_huge_zero_pmd(*pmd)) {
2810 __split_huge_zero_page_pmd(vma, haddr, pmd);
2811 spin_unlock(ptl);
2812 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2813 return;
2814 }
2815 page = pmd_page(*pmd);
2816 VM_BUG_ON_PAGE(!page_count(page), page);
2817 get_page(page);
2818 spin_unlock(ptl);
2819 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2820
2821 split_huge_page(page);
2822
2823 put_page(page);
2824
2825 /*
2826 * We don't always have down_write of mmap_sem here: a racing
2827 * do_huge_pmd_wp_page() might have copied-on-write to another
2828 * huge page before our split_huge_page() got the anon_vma lock.
2829 */
2830 if (unlikely(pmd_trans_huge(*pmd)))
2831 goto again;
2832}
2833
2834void split_huge_page_pmd_mm(struct mm_struct *mm, unsigned long address,
2835 pmd_t *pmd)
2836{
2837 struct vm_area_struct *vma;
2838
2839 vma = find_vma(mm, address);
2840 BUG_ON(vma == NULL);
2841 split_huge_page_pmd(vma, address, pmd);
2842}
2843
2844static void split_huge_page_address(struct mm_struct *mm,
2845 unsigned long address)
2846{
2847 pmd_t *pmd;
2848
2849 VM_BUG_ON(!(address & ~HPAGE_PMD_MASK));
2850
2851 pmd = mm_find_pmd(mm, address);
2852 if (!pmd)
2853 return;
2854 /*
2855 * Caller holds the mmap_sem write mode, so a huge pmd cannot
2856 * materialize from under us.
2857 */
2858 split_huge_page_pmd_mm(mm, address, pmd);
2859}
2860
2861void __vma_adjust_trans_huge(struct vm_area_struct *vma,
2862 unsigned long start,
2863 unsigned long end,
2864 long adjust_next)
2865{
2866 /*
2867 * If the new start address isn't hpage aligned and it could
2868 * previously contain an hugepage: check if we need to split
2869 * an huge pmd.
2870 */
2871 if (start & ~HPAGE_PMD_MASK &&
2872 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
2873 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2874 split_huge_page_address(vma->vm_mm, start);
2875
2876 /*
2877 * If the new end address isn't hpage aligned and it could
2878 * previously contain an hugepage: check if we need to split
2879 * an huge pmd.
2880 */
2881 if (end & ~HPAGE_PMD_MASK &&
2882 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
2883 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
2884 split_huge_page_address(vma->vm_mm, end);
2885
2886 /*
2887 * If we're also updating the vma->vm_next->vm_start, if the new
2888 * vm_next->vm_start isn't page aligned and it could previously
2889 * contain an hugepage: check if we need to split an huge pmd.
2890 */
2891 if (adjust_next > 0) {
2892 struct vm_area_struct *next = vma->vm_next;
2893 unsigned long nstart = next->vm_start;
2894 nstart += adjust_next << PAGE_SHIFT;
2895 if (nstart & ~HPAGE_PMD_MASK &&
2896 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
2897 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
2898 split_huge_page_address(next->vm_mm, nstart);
2899 }
2900}
1/*
2 * Copyright (C) 2009 Red Hat, Inc.
3 *
4 * This work is licensed under the terms of the GNU GPL, version 2. See
5 * the COPYING file in the top-level directory.
6 */
7
8#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
9
10#include <linux/mm.h>
11#include <linux/sched.h>
12#include <linux/highmem.h>
13#include <linux/hugetlb.h>
14#include <linux/mmu_notifier.h>
15#include <linux/rmap.h>
16#include <linux/swap.h>
17#include <linux/shrinker.h>
18#include <linux/mm_inline.h>
19#include <linux/swapops.h>
20#include <linux/dax.h>
21#include <linux/kthread.h>
22#include <linux/khugepaged.h>
23#include <linux/freezer.h>
24#include <linux/pfn_t.h>
25#include <linux/mman.h>
26#include <linux/memremap.h>
27#include <linux/pagemap.h>
28#include <linux/debugfs.h>
29#include <linux/migrate.h>
30#include <linux/hashtable.h>
31#include <linux/userfaultfd_k.h>
32#include <linux/page_idle.h>
33
34#include <asm/tlb.h>
35#include <asm/pgalloc.h>
36#include "internal.h"
37
38enum scan_result {
39 SCAN_FAIL,
40 SCAN_SUCCEED,
41 SCAN_PMD_NULL,
42 SCAN_EXCEED_NONE_PTE,
43 SCAN_PTE_NON_PRESENT,
44 SCAN_PAGE_RO,
45 SCAN_NO_REFERENCED_PAGE,
46 SCAN_PAGE_NULL,
47 SCAN_SCAN_ABORT,
48 SCAN_PAGE_COUNT,
49 SCAN_PAGE_LRU,
50 SCAN_PAGE_LOCK,
51 SCAN_PAGE_ANON,
52 SCAN_PAGE_COMPOUND,
53 SCAN_ANY_PROCESS,
54 SCAN_VMA_NULL,
55 SCAN_VMA_CHECK,
56 SCAN_ADDRESS_RANGE,
57 SCAN_SWAP_CACHE_PAGE,
58 SCAN_DEL_PAGE_LRU,
59 SCAN_ALLOC_HUGE_PAGE_FAIL,
60 SCAN_CGROUP_CHARGE_FAIL
61};
62
63#define CREATE_TRACE_POINTS
64#include <trace/events/huge_memory.h>
65
66/*
67 * By default transparent hugepage support is disabled in order that avoid
68 * to risk increase the memory footprint of applications without a guaranteed
69 * benefit. When transparent hugepage support is enabled, is for all mappings,
70 * and khugepaged scans all mappings.
71 * Defrag is invoked by khugepaged hugepage allocations and by page faults
72 * for all hugepage allocations.
73 */
74unsigned long transparent_hugepage_flags __read_mostly =
75#ifdef CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS
76 (1<<TRANSPARENT_HUGEPAGE_FLAG)|
77#endif
78#ifdef CONFIG_TRANSPARENT_HUGEPAGE_MADVISE
79 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)|
80#endif
81 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG)|
82 (1<<TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG)|
83 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
84
85/* default scan 8*512 pte (or vmas) every 30 second */
86static unsigned int khugepaged_pages_to_scan __read_mostly;
87static unsigned int khugepaged_pages_collapsed;
88static unsigned int khugepaged_full_scans;
89static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
90/* during fragmentation poll the hugepage allocator once every minute */
91static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
92static struct task_struct *khugepaged_thread __read_mostly;
93static DEFINE_MUTEX(khugepaged_mutex);
94static DEFINE_SPINLOCK(khugepaged_mm_lock);
95static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
96/*
97 * default collapse hugepages if there is at least one pte mapped like
98 * it would have happened if the vma was large enough during page
99 * fault.
100 */
101static unsigned int khugepaged_max_ptes_none __read_mostly;
102
103static int khugepaged(void *none);
104static int khugepaged_slab_init(void);
105static void khugepaged_slab_exit(void);
106
107#define MM_SLOTS_HASH_BITS 10
108static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
109
110static struct kmem_cache *mm_slot_cache __read_mostly;
111
112/**
113 * struct mm_slot - hash lookup from mm to mm_slot
114 * @hash: hash collision list
115 * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
116 * @mm: the mm that this information is valid for
117 */
118struct mm_slot {
119 struct hlist_node hash;
120 struct list_head mm_node;
121 struct mm_struct *mm;
122};
123
124/**
125 * struct khugepaged_scan - cursor for scanning
126 * @mm_head: the head of the mm list to scan
127 * @mm_slot: the current mm_slot we are scanning
128 * @address: the next address inside that to be scanned
129 *
130 * There is only the one khugepaged_scan instance of this cursor structure.
131 */
132struct khugepaged_scan {
133 struct list_head mm_head;
134 struct mm_slot *mm_slot;
135 unsigned long address;
136};
137static struct khugepaged_scan khugepaged_scan = {
138 .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
139};
140
141static struct shrinker deferred_split_shrinker;
142
143static void set_recommended_min_free_kbytes(void)
144{
145 struct zone *zone;
146 int nr_zones = 0;
147 unsigned long recommended_min;
148
149 for_each_populated_zone(zone)
150 nr_zones++;
151
152 /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
153 recommended_min = pageblock_nr_pages * nr_zones * 2;
154
155 /*
156 * Make sure that on average at least two pageblocks are almost free
157 * of another type, one for a migratetype to fall back to and a
158 * second to avoid subsequent fallbacks of other types There are 3
159 * MIGRATE_TYPES we care about.
160 */
161 recommended_min += pageblock_nr_pages * nr_zones *
162 MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
163
164 /* don't ever allow to reserve more than 5% of the lowmem */
165 recommended_min = min(recommended_min,
166 (unsigned long) nr_free_buffer_pages() / 20);
167 recommended_min <<= (PAGE_SHIFT-10);
168
169 if (recommended_min > min_free_kbytes) {
170 if (user_min_free_kbytes >= 0)
171 pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
172 min_free_kbytes, recommended_min);
173
174 min_free_kbytes = recommended_min;
175 }
176 setup_per_zone_wmarks();
177}
178
179static int start_stop_khugepaged(void)
180{
181 int err = 0;
182 if (khugepaged_enabled()) {
183 if (!khugepaged_thread)
184 khugepaged_thread = kthread_run(khugepaged, NULL,
185 "khugepaged");
186 if (IS_ERR(khugepaged_thread)) {
187 pr_err("khugepaged: kthread_run(khugepaged) failed\n");
188 err = PTR_ERR(khugepaged_thread);
189 khugepaged_thread = NULL;
190 goto fail;
191 }
192
193 if (!list_empty(&khugepaged_scan.mm_head))
194 wake_up_interruptible(&khugepaged_wait);
195
196 set_recommended_min_free_kbytes();
197 } else if (khugepaged_thread) {
198 kthread_stop(khugepaged_thread);
199 khugepaged_thread = NULL;
200 }
201fail:
202 return err;
203}
204
205static atomic_t huge_zero_refcount;
206struct page *huge_zero_page __read_mostly;
207
208struct page *get_huge_zero_page(void)
209{
210 struct page *zero_page;
211retry:
212 if (likely(atomic_inc_not_zero(&huge_zero_refcount)))
213 return READ_ONCE(huge_zero_page);
214
215 zero_page = alloc_pages((GFP_TRANSHUGE | __GFP_ZERO) & ~__GFP_MOVABLE,
216 HPAGE_PMD_ORDER);
217 if (!zero_page) {
218 count_vm_event(THP_ZERO_PAGE_ALLOC_FAILED);
219 return NULL;
220 }
221 count_vm_event(THP_ZERO_PAGE_ALLOC);
222 preempt_disable();
223 if (cmpxchg(&huge_zero_page, NULL, zero_page)) {
224 preempt_enable();
225 __free_pages(zero_page, compound_order(zero_page));
226 goto retry;
227 }
228
229 /* We take additional reference here. It will be put back by shrinker */
230 atomic_set(&huge_zero_refcount, 2);
231 preempt_enable();
232 return READ_ONCE(huge_zero_page);
233}
234
235void put_huge_zero_page(void)
236{
237 /*
238 * Counter should never go to zero here. Only shrinker can put
239 * last reference.
240 */
241 BUG_ON(atomic_dec_and_test(&huge_zero_refcount));
242}
243
244static unsigned long shrink_huge_zero_page_count(struct shrinker *shrink,
245 struct shrink_control *sc)
246{
247 /* we can free zero page only if last reference remains */
248 return atomic_read(&huge_zero_refcount) == 1 ? HPAGE_PMD_NR : 0;
249}
250
251static unsigned long shrink_huge_zero_page_scan(struct shrinker *shrink,
252 struct shrink_control *sc)
253{
254 if (atomic_cmpxchg(&huge_zero_refcount, 1, 0) == 1) {
255 struct page *zero_page = xchg(&huge_zero_page, NULL);
256 BUG_ON(zero_page == NULL);
257 __free_pages(zero_page, compound_order(zero_page));
258 return HPAGE_PMD_NR;
259 }
260
261 return 0;
262}
263
264static struct shrinker huge_zero_page_shrinker = {
265 .count_objects = shrink_huge_zero_page_count,
266 .scan_objects = shrink_huge_zero_page_scan,
267 .seeks = DEFAULT_SEEKS,
268};
269
270#ifdef CONFIG_SYSFS
271
272static ssize_t triple_flag_store(struct kobject *kobj,
273 struct kobj_attribute *attr,
274 const char *buf, size_t count,
275 enum transparent_hugepage_flag enabled,
276 enum transparent_hugepage_flag deferred,
277 enum transparent_hugepage_flag req_madv)
278{
279 if (!memcmp("defer", buf,
280 min(sizeof("defer")-1, count))) {
281 if (enabled == deferred)
282 return -EINVAL;
283 clear_bit(enabled, &transparent_hugepage_flags);
284 clear_bit(req_madv, &transparent_hugepage_flags);
285 set_bit(deferred, &transparent_hugepage_flags);
286 } else if (!memcmp("always", buf,
287 min(sizeof("always")-1, count))) {
288 clear_bit(deferred, &transparent_hugepage_flags);
289 clear_bit(req_madv, &transparent_hugepage_flags);
290 set_bit(enabled, &transparent_hugepage_flags);
291 } else if (!memcmp("madvise", buf,
292 min(sizeof("madvise")-1, count))) {
293 clear_bit(enabled, &transparent_hugepage_flags);
294 clear_bit(deferred, &transparent_hugepage_flags);
295 set_bit(req_madv, &transparent_hugepage_flags);
296 } else if (!memcmp("never", buf,
297 min(sizeof("never")-1, count))) {
298 clear_bit(enabled, &transparent_hugepage_flags);
299 clear_bit(req_madv, &transparent_hugepage_flags);
300 clear_bit(deferred, &transparent_hugepage_flags);
301 } else
302 return -EINVAL;
303
304 return count;
305}
306
307static ssize_t enabled_show(struct kobject *kobj,
308 struct kobj_attribute *attr, char *buf)
309{
310 if (test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags))
311 return sprintf(buf, "[always] madvise never\n");
312 else if (test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
313 return sprintf(buf, "always [madvise] never\n");
314 else
315 return sprintf(buf, "always madvise [never]\n");
316}
317
318static ssize_t enabled_store(struct kobject *kobj,
319 struct kobj_attribute *attr,
320 const char *buf, size_t count)
321{
322 ssize_t ret;
323
324 ret = triple_flag_store(kobj, attr, buf, count,
325 TRANSPARENT_HUGEPAGE_FLAG,
326 TRANSPARENT_HUGEPAGE_FLAG,
327 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG);
328
329 if (ret > 0) {
330 int err;
331
332 mutex_lock(&khugepaged_mutex);
333 err = start_stop_khugepaged();
334 mutex_unlock(&khugepaged_mutex);
335
336 if (err)
337 ret = err;
338 }
339
340 return ret;
341}
342static struct kobj_attribute enabled_attr =
343 __ATTR(enabled, 0644, enabled_show, enabled_store);
344
345static ssize_t single_flag_show(struct kobject *kobj,
346 struct kobj_attribute *attr, char *buf,
347 enum transparent_hugepage_flag flag)
348{
349 return sprintf(buf, "%d\n",
350 !!test_bit(flag, &transparent_hugepage_flags));
351}
352
353static ssize_t single_flag_store(struct kobject *kobj,
354 struct kobj_attribute *attr,
355 const char *buf, size_t count,
356 enum transparent_hugepage_flag flag)
357{
358 unsigned long value;
359 int ret;
360
361 ret = kstrtoul(buf, 10, &value);
362 if (ret < 0)
363 return ret;
364 if (value > 1)
365 return -EINVAL;
366
367 if (value)
368 set_bit(flag, &transparent_hugepage_flags);
369 else
370 clear_bit(flag, &transparent_hugepage_flags);
371
372 return count;
373}
374
375/*
376 * Currently defrag only disables __GFP_NOWAIT for allocation. A blind
377 * __GFP_REPEAT is too aggressive, it's never worth swapping tons of
378 * memory just to allocate one more hugepage.
379 */
380static ssize_t defrag_show(struct kobject *kobj,
381 struct kobj_attribute *attr, char *buf)
382{
383 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
384 return sprintf(buf, "[always] defer madvise never\n");
385 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
386 return sprintf(buf, "always [defer] madvise never\n");
387 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags))
388 return sprintf(buf, "always defer [madvise] never\n");
389 else
390 return sprintf(buf, "always defer madvise [never]\n");
391
392}
393static ssize_t defrag_store(struct kobject *kobj,
394 struct kobj_attribute *attr,
395 const char *buf, size_t count)
396{
397 return triple_flag_store(kobj, attr, buf, count,
398 TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
399 TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
400 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG);
401}
402static struct kobj_attribute defrag_attr =
403 __ATTR(defrag, 0644, defrag_show, defrag_store);
404
405static ssize_t use_zero_page_show(struct kobject *kobj,
406 struct kobj_attribute *attr, char *buf)
407{
408 return single_flag_show(kobj, attr, buf,
409 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
410}
411static ssize_t use_zero_page_store(struct kobject *kobj,
412 struct kobj_attribute *attr, const char *buf, size_t count)
413{
414 return single_flag_store(kobj, attr, buf, count,
415 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG);
416}
417static struct kobj_attribute use_zero_page_attr =
418 __ATTR(use_zero_page, 0644, use_zero_page_show, use_zero_page_store);
419#ifdef CONFIG_DEBUG_VM
420static ssize_t debug_cow_show(struct kobject *kobj,
421 struct kobj_attribute *attr, char *buf)
422{
423 return single_flag_show(kobj, attr, buf,
424 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
425}
426static ssize_t debug_cow_store(struct kobject *kobj,
427 struct kobj_attribute *attr,
428 const char *buf, size_t count)
429{
430 return single_flag_store(kobj, attr, buf, count,
431 TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG);
432}
433static struct kobj_attribute debug_cow_attr =
434 __ATTR(debug_cow, 0644, debug_cow_show, debug_cow_store);
435#endif /* CONFIG_DEBUG_VM */
436
437static struct attribute *hugepage_attr[] = {
438 &enabled_attr.attr,
439 &defrag_attr.attr,
440 &use_zero_page_attr.attr,
441#ifdef CONFIG_DEBUG_VM
442 &debug_cow_attr.attr,
443#endif
444 NULL,
445};
446
447static struct attribute_group hugepage_attr_group = {
448 .attrs = hugepage_attr,
449};
450
451static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
452 struct kobj_attribute *attr,
453 char *buf)
454{
455 return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
456}
457
458static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
459 struct kobj_attribute *attr,
460 const char *buf, size_t count)
461{
462 unsigned long msecs;
463 int err;
464
465 err = kstrtoul(buf, 10, &msecs);
466 if (err || msecs > UINT_MAX)
467 return -EINVAL;
468
469 khugepaged_scan_sleep_millisecs = msecs;
470 wake_up_interruptible(&khugepaged_wait);
471
472 return count;
473}
474static struct kobj_attribute scan_sleep_millisecs_attr =
475 __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
476 scan_sleep_millisecs_store);
477
478static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
479 struct kobj_attribute *attr,
480 char *buf)
481{
482 return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
483}
484
485static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
486 struct kobj_attribute *attr,
487 const char *buf, size_t count)
488{
489 unsigned long msecs;
490 int err;
491
492 err = kstrtoul(buf, 10, &msecs);
493 if (err || msecs > UINT_MAX)
494 return -EINVAL;
495
496 khugepaged_alloc_sleep_millisecs = msecs;
497 wake_up_interruptible(&khugepaged_wait);
498
499 return count;
500}
501static struct kobj_attribute alloc_sleep_millisecs_attr =
502 __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
503 alloc_sleep_millisecs_store);
504
505static ssize_t pages_to_scan_show(struct kobject *kobj,
506 struct kobj_attribute *attr,
507 char *buf)
508{
509 return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
510}
511static ssize_t pages_to_scan_store(struct kobject *kobj,
512 struct kobj_attribute *attr,
513 const char *buf, size_t count)
514{
515 int err;
516 unsigned long pages;
517
518 err = kstrtoul(buf, 10, &pages);
519 if (err || !pages || pages > UINT_MAX)
520 return -EINVAL;
521
522 khugepaged_pages_to_scan = pages;
523
524 return count;
525}
526static struct kobj_attribute pages_to_scan_attr =
527 __ATTR(pages_to_scan, 0644, pages_to_scan_show,
528 pages_to_scan_store);
529
530static ssize_t pages_collapsed_show(struct kobject *kobj,
531 struct kobj_attribute *attr,
532 char *buf)
533{
534 return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
535}
536static struct kobj_attribute pages_collapsed_attr =
537 __ATTR_RO(pages_collapsed);
538
539static ssize_t full_scans_show(struct kobject *kobj,
540 struct kobj_attribute *attr,
541 char *buf)
542{
543 return sprintf(buf, "%u\n", khugepaged_full_scans);
544}
545static struct kobj_attribute full_scans_attr =
546 __ATTR_RO(full_scans);
547
548static ssize_t khugepaged_defrag_show(struct kobject *kobj,
549 struct kobj_attribute *attr, char *buf)
550{
551 return single_flag_show(kobj, attr, buf,
552 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
553}
554static ssize_t khugepaged_defrag_store(struct kobject *kobj,
555 struct kobj_attribute *attr,
556 const char *buf, size_t count)
557{
558 return single_flag_store(kobj, attr, buf, count,
559 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
560}
561static struct kobj_attribute khugepaged_defrag_attr =
562 __ATTR(defrag, 0644, khugepaged_defrag_show,
563 khugepaged_defrag_store);
564
565/*
566 * max_ptes_none controls if khugepaged should collapse hugepages over
567 * any unmapped ptes in turn potentially increasing the memory
568 * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
569 * reduce the available free memory in the system as it
570 * runs. Increasing max_ptes_none will instead potentially reduce the
571 * free memory in the system during the khugepaged scan.
572 */
573static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
574 struct kobj_attribute *attr,
575 char *buf)
576{
577 return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
578}
579static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
580 struct kobj_attribute *attr,
581 const char *buf, size_t count)
582{
583 int err;
584 unsigned long max_ptes_none;
585
586 err = kstrtoul(buf, 10, &max_ptes_none);
587 if (err || max_ptes_none > HPAGE_PMD_NR-1)
588 return -EINVAL;
589
590 khugepaged_max_ptes_none = max_ptes_none;
591
592 return count;
593}
594static struct kobj_attribute khugepaged_max_ptes_none_attr =
595 __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
596 khugepaged_max_ptes_none_store);
597
598static struct attribute *khugepaged_attr[] = {
599 &khugepaged_defrag_attr.attr,
600 &khugepaged_max_ptes_none_attr.attr,
601 &pages_to_scan_attr.attr,
602 &pages_collapsed_attr.attr,
603 &full_scans_attr.attr,
604 &scan_sleep_millisecs_attr.attr,
605 &alloc_sleep_millisecs_attr.attr,
606 NULL,
607};
608
609static struct attribute_group khugepaged_attr_group = {
610 .attrs = khugepaged_attr,
611 .name = "khugepaged",
612};
613
614static int __init hugepage_init_sysfs(struct kobject **hugepage_kobj)
615{
616 int err;
617
618 *hugepage_kobj = kobject_create_and_add("transparent_hugepage", mm_kobj);
619 if (unlikely(!*hugepage_kobj)) {
620 pr_err("failed to create transparent hugepage kobject\n");
621 return -ENOMEM;
622 }
623
624 err = sysfs_create_group(*hugepage_kobj, &hugepage_attr_group);
625 if (err) {
626 pr_err("failed to register transparent hugepage group\n");
627 goto delete_obj;
628 }
629
630 err = sysfs_create_group(*hugepage_kobj, &khugepaged_attr_group);
631 if (err) {
632 pr_err("failed to register transparent hugepage group\n");
633 goto remove_hp_group;
634 }
635
636 return 0;
637
638remove_hp_group:
639 sysfs_remove_group(*hugepage_kobj, &hugepage_attr_group);
640delete_obj:
641 kobject_put(*hugepage_kobj);
642 return err;
643}
644
645static void __init hugepage_exit_sysfs(struct kobject *hugepage_kobj)
646{
647 sysfs_remove_group(hugepage_kobj, &khugepaged_attr_group);
648 sysfs_remove_group(hugepage_kobj, &hugepage_attr_group);
649 kobject_put(hugepage_kobj);
650}
651#else
652static inline int hugepage_init_sysfs(struct kobject **hugepage_kobj)
653{
654 return 0;
655}
656
657static inline void hugepage_exit_sysfs(struct kobject *hugepage_kobj)
658{
659}
660#endif /* CONFIG_SYSFS */
661
662static int __init hugepage_init(void)
663{
664 int err;
665 struct kobject *hugepage_kobj;
666
667 if (!has_transparent_hugepage()) {
668 transparent_hugepage_flags = 0;
669 return -EINVAL;
670 }
671
672 khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
673 khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
674 /*
675 * hugepages can't be allocated by the buddy allocator
676 */
677 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER >= MAX_ORDER);
678 /*
679 * we use page->mapping and page->index in second tail page
680 * as list_head: assuming THP order >= 2
681 */
682 MAYBE_BUILD_BUG_ON(HPAGE_PMD_ORDER < 2);
683
684 err = hugepage_init_sysfs(&hugepage_kobj);
685 if (err)
686 goto err_sysfs;
687
688 err = khugepaged_slab_init();
689 if (err)
690 goto err_slab;
691
692 err = register_shrinker(&huge_zero_page_shrinker);
693 if (err)
694 goto err_hzp_shrinker;
695 err = register_shrinker(&deferred_split_shrinker);
696 if (err)
697 goto err_split_shrinker;
698
699 /*
700 * By default disable transparent hugepages on smaller systems,
701 * where the extra memory used could hurt more than TLB overhead
702 * is likely to save. The admin can still enable it through /sys.
703 */
704 if (totalram_pages < (512 << (20 - PAGE_SHIFT))) {
705 transparent_hugepage_flags = 0;
706 return 0;
707 }
708
709 err = start_stop_khugepaged();
710 if (err)
711 goto err_khugepaged;
712
713 return 0;
714err_khugepaged:
715 unregister_shrinker(&deferred_split_shrinker);
716err_split_shrinker:
717 unregister_shrinker(&huge_zero_page_shrinker);
718err_hzp_shrinker:
719 khugepaged_slab_exit();
720err_slab:
721 hugepage_exit_sysfs(hugepage_kobj);
722err_sysfs:
723 return err;
724}
725subsys_initcall(hugepage_init);
726
727static int __init setup_transparent_hugepage(char *str)
728{
729 int ret = 0;
730 if (!str)
731 goto out;
732 if (!strcmp(str, "always")) {
733 set_bit(TRANSPARENT_HUGEPAGE_FLAG,
734 &transparent_hugepage_flags);
735 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
736 &transparent_hugepage_flags);
737 ret = 1;
738 } else if (!strcmp(str, "madvise")) {
739 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
740 &transparent_hugepage_flags);
741 set_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
742 &transparent_hugepage_flags);
743 ret = 1;
744 } else if (!strcmp(str, "never")) {
745 clear_bit(TRANSPARENT_HUGEPAGE_FLAG,
746 &transparent_hugepage_flags);
747 clear_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
748 &transparent_hugepage_flags);
749 ret = 1;
750 }
751out:
752 if (!ret)
753 pr_warn("transparent_hugepage= cannot parse, ignored\n");
754 return ret;
755}
756__setup("transparent_hugepage=", setup_transparent_hugepage);
757
758pmd_t maybe_pmd_mkwrite(pmd_t pmd, struct vm_area_struct *vma)
759{
760 if (likely(vma->vm_flags & VM_WRITE))
761 pmd = pmd_mkwrite(pmd);
762 return pmd;
763}
764
765static inline pmd_t mk_huge_pmd(struct page *page, pgprot_t prot)
766{
767 pmd_t entry;
768 entry = mk_pmd(page, prot);
769 entry = pmd_mkhuge(entry);
770 return entry;
771}
772
773static inline struct list_head *page_deferred_list(struct page *page)
774{
775 /*
776 * ->lru in the tail pages is occupied by compound_head.
777 * Let's use ->mapping + ->index in the second tail page as list_head.
778 */
779 return (struct list_head *)&page[2].mapping;
780}
781
782void prep_transhuge_page(struct page *page)
783{
784 /*
785 * we use page->mapping and page->indexlru in second tail page
786 * as list_head: assuming THP order >= 2
787 */
788
789 INIT_LIST_HEAD(page_deferred_list(page));
790 set_compound_page_dtor(page, TRANSHUGE_PAGE_DTOR);
791}
792
793static int __do_huge_pmd_anonymous_page(struct mm_struct *mm,
794 struct vm_area_struct *vma,
795 unsigned long address, pmd_t *pmd,
796 struct page *page, gfp_t gfp,
797 unsigned int flags)
798{
799 struct mem_cgroup *memcg;
800 pgtable_t pgtable;
801 spinlock_t *ptl;
802 unsigned long haddr = address & HPAGE_PMD_MASK;
803
804 VM_BUG_ON_PAGE(!PageCompound(page), page);
805
806 if (mem_cgroup_try_charge(page, mm, gfp, &memcg, true)) {
807 put_page(page);
808 count_vm_event(THP_FAULT_FALLBACK);
809 return VM_FAULT_FALLBACK;
810 }
811
812 pgtable = pte_alloc_one(mm, haddr);
813 if (unlikely(!pgtable)) {
814 mem_cgroup_cancel_charge(page, memcg, true);
815 put_page(page);
816 return VM_FAULT_OOM;
817 }
818
819 clear_huge_page(page, haddr, HPAGE_PMD_NR);
820 /*
821 * The memory barrier inside __SetPageUptodate makes sure that
822 * clear_huge_page writes become visible before the set_pmd_at()
823 * write.
824 */
825 __SetPageUptodate(page);
826
827 ptl = pmd_lock(mm, pmd);
828 if (unlikely(!pmd_none(*pmd))) {
829 spin_unlock(ptl);
830 mem_cgroup_cancel_charge(page, memcg, true);
831 put_page(page);
832 pte_free(mm, pgtable);
833 } else {
834 pmd_t entry;
835
836 /* Deliver the page fault to userland */
837 if (userfaultfd_missing(vma)) {
838 int ret;
839
840 spin_unlock(ptl);
841 mem_cgroup_cancel_charge(page, memcg, true);
842 put_page(page);
843 pte_free(mm, pgtable);
844 ret = handle_userfault(vma, address, flags,
845 VM_UFFD_MISSING);
846 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
847 return ret;
848 }
849
850 entry = mk_huge_pmd(page, vma->vm_page_prot);
851 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
852 page_add_new_anon_rmap(page, vma, haddr, true);
853 mem_cgroup_commit_charge(page, memcg, false, true);
854 lru_cache_add_active_or_unevictable(page, vma);
855 pgtable_trans_huge_deposit(mm, pmd, pgtable);
856 set_pmd_at(mm, haddr, pmd, entry);
857 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
858 atomic_long_inc(&mm->nr_ptes);
859 spin_unlock(ptl);
860 count_vm_event(THP_FAULT_ALLOC);
861 }
862
863 return 0;
864}
865
866/*
867 * If THP is set to always then directly reclaim/compact as necessary
868 * If set to defer then do no reclaim and defer to khugepaged
869 * If set to madvise and the VMA is flagged then directly reclaim/compact
870 */
871static inline gfp_t alloc_hugepage_direct_gfpmask(struct vm_area_struct *vma)
872{
873 gfp_t reclaim_flags = 0;
874
875 if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG, &transparent_hugepage_flags) &&
876 (vma->vm_flags & VM_HUGEPAGE))
877 reclaim_flags = __GFP_DIRECT_RECLAIM;
878 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags))
879 reclaim_flags = __GFP_KSWAPD_RECLAIM;
880 else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG, &transparent_hugepage_flags))
881 reclaim_flags = __GFP_DIRECT_RECLAIM;
882
883 return GFP_TRANSHUGE | reclaim_flags;
884}
885
886/* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
887static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
888{
889 return GFP_TRANSHUGE | (khugepaged_defrag() ? __GFP_DIRECT_RECLAIM : 0);
890}
891
892/* Caller must hold page table lock. */
893static bool set_huge_zero_page(pgtable_t pgtable, struct mm_struct *mm,
894 struct vm_area_struct *vma, unsigned long haddr, pmd_t *pmd,
895 struct page *zero_page)
896{
897 pmd_t entry;
898 if (!pmd_none(*pmd))
899 return false;
900 entry = mk_pmd(zero_page, vma->vm_page_prot);
901 entry = pmd_mkhuge(entry);
902 if (pgtable)
903 pgtable_trans_huge_deposit(mm, pmd, pgtable);
904 set_pmd_at(mm, haddr, pmd, entry);
905 atomic_long_inc(&mm->nr_ptes);
906 return true;
907}
908
909int do_huge_pmd_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
910 unsigned long address, pmd_t *pmd,
911 unsigned int flags)
912{
913 gfp_t gfp;
914 struct page *page;
915 unsigned long haddr = address & HPAGE_PMD_MASK;
916
917 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
918 return VM_FAULT_FALLBACK;
919 if (unlikely(anon_vma_prepare(vma)))
920 return VM_FAULT_OOM;
921 if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
922 return VM_FAULT_OOM;
923 if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm) &&
924 transparent_hugepage_use_zero_page()) {
925 spinlock_t *ptl;
926 pgtable_t pgtable;
927 struct page *zero_page;
928 bool set;
929 int ret;
930 pgtable = pte_alloc_one(mm, haddr);
931 if (unlikely(!pgtable))
932 return VM_FAULT_OOM;
933 zero_page = get_huge_zero_page();
934 if (unlikely(!zero_page)) {
935 pte_free(mm, pgtable);
936 count_vm_event(THP_FAULT_FALLBACK);
937 return VM_FAULT_FALLBACK;
938 }
939 ptl = pmd_lock(mm, pmd);
940 ret = 0;
941 set = false;
942 if (pmd_none(*pmd)) {
943 if (userfaultfd_missing(vma)) {
944 spin_unlock(ptl);
945 ret = handle_userfault(vma, address, flags,
946 VM_UFFD_MISSING);
947 VM_BUG_ON(ret & VM_FAULT_FALLBACK);
948 } else {
949 set_huge_zero_page(pgtable, mm, vma,
950 haddr, pmd,
951 zero_page);
952 spin_unlock(ptl);
953 set = true;
954 }
955 } else
956 spin_unlock(ptl);
957 if (!set) {
958 pte_free(mm, pgtable);
959 put_huge_zero_page();
960 }
961 return ret;
962 }
963 gfp = alloc_hugepage_direct_gfpmask(vma);
964 page = alloc_hugepage_vma(gfp, vma, haddr, HPAGE_PMD_ORDER);
965 if (unlikely(!page)) {
966 count_vm_event(THP_FAULT_FALLBACK);
967 return VM_FAULT_FALLBACK;
968 }
969 prep_transhuge_page(page);
970 return __do_huge_pmd_anonymous_page(mm, vma, address, pmd, page, gfp,
971 flags);
972}
973
974static void insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
975 pmd_t *pmd, pfn_t pfn, pgprot_t prot, bool write)
976{
977 struct mm_struct *mm = vma->vm_mm;
978 pmd_t entry;
979 spinlock_t *ptl;
980
981 ptl = pmd_lock(mm, pmd);
982 entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
983 if (pfn_t_devmap(pfn))
984 entry = pmd_mkdevmap(entry);
985 if (write) {
986 entry = pmd_mkyoung(pmd_mkdirty(entry));
987 entry = maybe_pmd_mkwrite(entry, vma);
988 }
989 set_pmd_at(mm, addr, pmd, entry);
990 update_mmu_cache_pmd(vma, addr, pmd);
991 spin_unlock(ptl);
992}
993
994int vmf_insert_pfn_pmd(struct vm_area_struct *vma, unsigned long addr,
995 pmd_t *pmd, pfn_t pfn, bool write)
996{
997 pgprot_t pgprot = vma->vm_page_prot;
998 /*
999 * If we had pmd_special, we could avoid all these restrictions,
1000 * but we need to be consistent with PTEs and architectures that
1001 * can't support a 'special' bit.
1002 */
1003 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1004 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1005 (VM_PFNMAP|VM_MIXEDMAP));
1006 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1007 BUG_ON(!pfn_t_devmap(pfn));
1008
1009 if (addr < vma->vm_start || addr >= vma->vm_end)
1010 return VM_FAULT_SIGBUS;
1011 if (track_pfn_insert(vma, &pgprot, pfn))
1012 return VM_FAULT_SIGBUS;
1013 insert_pfn_pmd(vma, addr, pmd, pfn, pgprot, write);
1014 return VM_FAULT_NOPAGE;
1015}
1016
1017static void touch_pmd(struct vm_area_struct *vma, unsigned long addr,
1018 pmd_t *pmd)
1019{
1020 pmd_t _pmd;
1021
1022 /*
1023 * We should set the dirty bit only for FOLL_WRITE but for now
1024 * the dirty bit in the pmd is meaningless. And if the dirty
1025 * bit will become meaningful and we'll only set it with
1026 * FOLL_WRITE, an atomic set_bit will be required on the pmd to
1027 * set the young bit, instead of the current set_pmd_at.
1028 */
1029 _pmd = pmd_mkyoung(pmd_mkdirty(*pmd));
1030 if (pmdp_set_access_flags(vma, addr & HPAGE_PMD_MASK,
1031 pmd, _pmd, 1))
1032 update_mmu_cache_pmd(vma, addr, pmd);
1033}
1034
1035struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
1036 pmd_t *pmd, int flags)
1037{
1038 unsigned long pfn = pmd_pfn(*pmd);
1039 struct mm_struct *mm = vma->vm_mm;
1040 struct dev_pagemap *pgmap;
1041 struct page *page;
1042
1043 assert_spin_locked(pmd_lockptr(mm, pmd));
1044
1045 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1046 return NULL;
1047
1048 if (pmd_present(*pmd) && pmd_devmap(*pmd))
1049 /* pass */;
1050 else
1051 return NULL;
1052
1053 if (flags & FOLL_TOUCH)
1054 touch_pmd(vma, addr, pmd);
1055
1056 /*
1057 * device mapped pages can only be returned if the
1058 * caller will manage the page reference count.
1059 */
1060 if (!(flags & FOLL_GET))
1061 return ERR_PTR(-EEXIST);
1062
1063 pfn += (addr & ~PMD_MASK) >> PAGE_SHIFT;
1064 pgmap = get_dev_pagemap(pfn, NULL);
1065 if (!pgmap)
1066 return ERR_PTR(-EFAULT);
1067 page = pfn_to_page(pfn);
1068 get_page(page);
1069 put_dev_pagemap(pgmap);
1070
1071 return page;
1072}
1073
1074int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1075 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
1076 struct vm_area_struct *vma)
1077{
1078 spinlock_t *dst_ptl, *src_ptl;
1079 struct page *src_page;
1080 pmd_t pmd;
1081 pgtable_t pgtable = NULL;
1082 int ret;
1083
1084 if (!vma_is_dax(vma)) {
1085 ret = -ENOMEM;
1086 pgtable = pte_alloc_one(dst_mm, addr);
1087 if (unlikely(!pgtable))
1088 goto out;
1089 }
1090
1091 dst_ptl = pmd_lock(dst_mm, dst_pmd);
1092 src_ptl = pmd_lockptr(src_mm, src_pmd);
1093 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
1094
1095 ret = -EAGAIN;
1096 pmd = *src_pmd;
1097 if (unlikely(!pmd_trans_huge(pmd) && !pmd_devmap(pmd))) {
1098 pte_free(dst_mm, pgtable);
1099 goto out_unlock;
1100 }
1101 /*
1102 * When page table lock is held, the huge zero pmd should not be
1103 * under splitting since we don't split the page itself, only pmd to
1104 * a page table.
1105 */
1106 if (is_huge_zero_pmd(pmd)) {
1107 struct page *zero_page;
1108 /*
1109 * get_huge_zero_page() will never allocate a new page here,
1110 * since we already have a zero page to copy. It just takes a
1111 * reference.
1112 */
1113 zero_page = get_huge_zero_page();
1114 set_huge_zero_page(pgtable, dst_mm, vma, addr, dst_pmd,
1115 zero_page);
1116 ret = 0;
1117 goto out_unlock;
1118 }
1119
1120 if (!vma_is_dax(vma)) {
1121 /* thp accounting separate from pmd_devmap accounting */
1122 src_page = pmd_page(pmd);
1123 VM_BUG_ON_PAGE(!PageHead(src_page), src_page);
1124 get_page(src_page);
1125 page_dup_rmap(src_page, true);
1126 add_mm_counter(dst_mm, MM_ANONPAGES, HPAGE_PMD_NR);
1127 atomic_long_inc(&dst_mm->nr_ptes);
1128 pgtable_trans_huge_deposit(dst_mm, dst_pmd, pgtable);
1129 }
1130
1131 pmdp_set_wrprotect(src_mm, addr, src_pmd);
1132 pmd = pmd_mkold(pmd_wrprotect(pmd));
1133 set_pmd_at(dst_mm, addr, dst_pmd, pmd);
1134
1135 ret = 0;
1136out_unlock:
1137 spin_unlock(src_ptl);
1138 spin_unlock(dst_ptl);
1139out:
1140 return ret;
1141}
1142
1143void huge_pmd_set_accessed(struct mm_struct *mm,
1144 struct vm_area_struct *vma,
1145 unsigned long address,
1146 pmd_t *pmd, pmd_t orig_pmd,
1147 int dirty)
1148{
1149 spinlock_t *ptl;
1150 pmd_t entry;
1151 unsigned long haddr;
1152
1153 ptl = pmd_lock(mm, pmd);
1154 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1155 goto unlock;
1156
1157 entry = pmd_mkyoung(orig_pmd);
1158 haddr = address & HPAGE_PMD_MASK;
1159 if (pmdp_set_access_flags(vma, haddr, pmd, entry, dirty))
1160 update_mmu_cache_pmd(vma, address, pmd);
1161
1162unlock:
1163 spin_unlock(ptl);
1164}
1165
1166static int do_huge_pmd_wp_page_fallback(struct mm_struct *mm,
1167 struct vm_area_struct *vma,
1168 unsigned long address,
1169 pmd_t *pmd, pmd_t orig_pmd,
1170 struct page *page,
1171 unsigned long haddr)
1172{
1173 struct mem_cgroup *memcg;
1174 spinlock_t *ptl;
1175 pgtable_t pgtable;
1176 pmd_t _pmd;
1177 int ret = 0, i;
1178 struct page **pages;
1179 unsigned long mmun_start; /* For mmu_notifiers */
1180 unsigned long mmun_end; /* For mmu_notifiers */
1181
1182 pages = kmalloc(sizeof(struct page *) * HPAGE_PMD_NR,
1183 GFP_KERNEL);
1184 if (unlikely(!pages)) {
1185 ret |= VM_FAULT_OOM;
1186 goto out;
1187 }
1188
1189 for (i = 0; i < HPAGE_PMD_NR; i++) {
1190 pages[i] = alloc_page_vma_node(GFP_HIGHUSER_MOVABLE |
1191 __GFP_OTHER_NODE,
1192 vma, address, page_to_nid(page));
1193 if (unlikely(!pages[i] ||
1194 mem_cgroup_try_charge(pages[i], mm, GFP_KERNEL,
1195 &memcg, false))) {
1196 if (pages[i])
1197 put_page(pages[i]);
1198 while (--i >= 0) {
1199 memcg = (void *)page_private(pages[i]);
1200 set_page_private(pages[i], 0);
1201 mem_cgroup_cancel_charge(pages[i], memcg,
1202 false);
1203 put_page(pages[i]);
1204 }
1205 kfree(pages);
1206 ret |= VM_FAULT_OOM;
1207 goto out;
1208 }
1209 set_page_private(pages[i], (unsigned long)memcg);
1210 }
1211
1212 for (i = 0; i < HPAGE_PMD_NR; i++) {
1213 copy_user_highpage(pages[i], page + i,
1214 haddr + PAGE_SIZE * i, vma);
1215 __SetPageUptodate(pages[i]);
1216 cond_resched();
1217 }
1218
1219 mmun_start = haddr;
1220 mmun_end = haddr + HPAGE_PMD_SIZE;
1221 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1222
1223 ptl = pmd_lock(mm, pmd);
1224 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1225 goto out_free_pages;
1226 VM_BUG_ON_PAGE(!PageHead(page), page);
1227
1228 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1229 /* leave pmd empty until pte is filled */
1230
1231 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
1232 pmd_populate(mm, &_pmd, pgtable);
1233
1234 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
1235 pte_t *pte, entry;
1236 entry = mk_pte(pages[i], vma->vm_page_prot);
1237 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1238 memcg = (void *)page_private(pages[i]);
1239 set_page_private(pages[i], 0);
1240 page_add_new_anon_rmap(pages[i], vma, haddr, false);
1241 mem_cgroup_commit_charge(pages[i], memcg, false, false);
1242 lru_cache_add_active_or_unevictable(pages[i], vma);
1243 pte = pte_offset_map(&_pmd, haddr);
1244 VM_BUG_ON(!pte_none(*pte));
1245 set_pte_at(mm, haddr, pte, entry);
1246 pte_unmap(pte);
1247 }
1248 kfree(pages);
1249
1250 smp_wmb(); /* make pte visible before pmd */
1251 pmd_populate(mm, pmd, pgtable);
1252 page_remove_rmap(page, true);
1253 spin_unlock(ptl);
1254
1255 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1256
1257 ret |= VM_FAULT_WRITE;
1258 put_page(page);
1259
1260out:
1261 return ret;
1262
1263out_free_pages:
1264 spin_unlock(ptl);
1265 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1266 for (i = 0; i < HPAGE_PMD_NR; i++) {
1267 memcg = (void *)page_private(pages[i]);
1268 set_page_private(pages[i], 0);
1269 mem_cgroup_cancel_charge(pages[i], memcg, false);
1270 put_page(pages[i]);
1271 }
1272 kfree(pages);
1273 goto out;
1274}
1275
1276int do_huge_pmd_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1277 unsigned long address, pmd_t *pmd, pmd_t orig_pmd)
1278{
1279 spinlock_t *ptl;
1280 int ret = 0;
1281 struct page *page = NULL, *new_page;
1282 struct mem_cgroup *memcg;
1283 unsigned long haddr;
1284 unsigned long mmun_start; /* For mmu_notifiers */
1285 unsigned long mmun_end; /* For mmu_notifiers */
1286 gfp_t huge_gfp; /* for allocation and charge */
1287
1288 ptl = pmd_lockptr(mm, pmd);
1289 VM_BUG_ON_VMA(!vma->anon_vma, vma);
1290 haddr = address & HPAGE_PMD_MASK;
1291 if (is_huge_zero_pmd(orig_pmd))
1292 goto alloc;
1293 spin_lock(ptl);
1294 if (unlikely(!pmd_same(*pmd, orig_pmd)))
1295 goto out_unlock;
1296
1297 page = pmd_page(orig_pmd);
1298 VM_BUG_ON_PAGE(!PageCompound(page) || !PageHead(page), page);
1299 /*
1300 * We can only reuse the page if nobody else maps the huge page or it's
1301 * part.
1302 */
1303 if (page_trans_huge_mapcount(page, NULL) == 1) {
1304 pmd_t entry;
1305 entry = pmd_mkyoung(orig_pmd);
1306 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1307 if (pmdp_set_access_flags(vma, haddr, pmd, entry, 1))
1308 update_mmu_cache_pmd(vma, address, pmd);
1309 ret |= VM_FAULT_WRITE;
1310 goto out_unlock;
1311 }
1312 get_page(page);
1313 spin_unlock(ptl);
1314alloc:
1315 if (transparent_hugepage_enabled(vma) &&
1316 !transparent_hugepage_debug_cow()) {
1317 huge_gfp = alloc_hugepage_direct_gfpmask(vma);
1318 new_page = alloc_hugepage_vma(huge_gfp, vma, haddr, HPAGE_PMD_ORDER);
1319 } else
1320 new_page = NULL;
1321
1322 if (likely(new_page)) {
1323 prep_transhuge_page(new_page);
1324 } else {
1325 if (!page) {
1326 split_huge_pmd(vma, pmd, address);
1327 ret |= VM_FAULT_FALLBACK;
1328 } else {
1329 ret = do_huge_pmd_wp_page_fallback(mm, vma, address,
1330 pmd, orig_pmd, page, haddr);
1331 if (ret & VM_FAULT_OOM) {
1332 split_huge_pmd(vma, pmd, address);
1333 ret |= VM_FAULT_FALLBACK;
1334 }
1335 put_page(page);
1336 }
1337 count_vm_event(THP_FAULT_FALLBACK);
1338 goto out;
1339 }
1340
1341 if (unlikely(mem_cgroup_try_charge(new_page, mm, huge_gfp, &memcg,
1342 true))) {
1343 put_page(new_page);
1344 if (page) {
1345 split_huge_pmd(vma, pmd, address);
1346 put_page(page);
1347 } else
1348 split_huge_pmd(vma, pmd, address);
1349 ret |= VM_FAULT_FALLBACK;
1350 count_vm_event(THP_FAULT_FALLBACK);
1351 goto out;
1352 }
1353
1354 count_vm_event(THP_FAULT_ALLOC);
1355
1356 if (!page)
1357 clear_huge_page(new_page, haddr, HPAGE_PMD_NR);
1358 else
1359 copy_user_huge_page(new_page, page, haddr, vma, HPAGE_PMD_NR);
1360 __SetPageUptodate(new_page);
1361
1362 mmun_start = haddr;
1363 mmun_end = haddr + HPAGE_PMD_SIZE;
1364 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1365
1366 spin_lock(ptl);
1367 if (page)
1368 put_page(page);
1369 if (unlikely(!pmd_same(*pmd, orig_pmd))) {
1370 spin_unlock(ptl);
1371 mem_cgroup_cancel_charge(new_page, memcg, true);
1372 put_page(new_page);
1373 goto out_mn;
1374 } else {
1375 pmd_t entry;
1376 entry = mk_huge_pmd(new_page, vma->vm_page_prot);
1377 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
1378 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
1379 page_add_new_anon_rmap(new_page, vma, haddr, true);
1380 mem_cgroup_commit_charge(new_page, memcg, false, true);
1381 lru_cache_add_active_or_unevictable(new_page, vma);
1382 set_pmd_at(mm, haddr, pmd, entry);
1383 update_mmu_cache_pmd(vma, address, pmd);
1384 if (!page) {
1385 add_mm_counter(mm, MM_ANONPAGES, HPAGE_PMD_NR);
1386 put_huge_zero_page();
1387 } else {
1388 VM_BUG_ON_PAGE(!PageHead(page), page);
1389 page_remove_rmap(page, true);
1390 put_page(page);
1391 }
1392 ret |= VM_FAULT_WRITE;
1393 }
1394 spin_unlock(ptl);
1395out_mn:
1396 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1397out:
1398 return ret;
1399out_unlock:
1400 spin_unlock(ptl);
1401 return ret;
1402}
1403
1404struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
1405 unsigned long addr,
1406 pmd_t *pmd,
1407 unsigned int flags)
1408{
1409 struct mm_struct *mm = vma->vm_mm;
1410 struct page *page = NULL;
1411
1412 assert_spin_locked(pmd_lockptr(mm, pmd));
1413
1414 if (flags & FOLL_WRITE && !pmd_write(*pmd))
1415 goto out;
1416
1417 /* Avoid dumping huge zero page */
1418 if ((flags & FOLL_DUMP) && is_huge_zero_pmd(*pmd))
1419 return ERR_PTR(-EFAULT);
1420
1421 /* Full NUMA hinting faults to serialise migration in fault paths */
1422 if ((flags & FOLL_NUMA) && pmd_protnone(*pmd))
1423 goto out;
1424
1425 page = pmd_page(*pmd);
1426 VM_BUG_ON_PAGE(!PageHead(page), page);
1427 if (flags & FOLL_TOUCH)
1428 touch_pmd(vma, addr, pmd);
1429 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1430 /*
1431 * We don't mlock() pte-mapped THPs. This way we can avoid
1432 * leaking mlocked pages into non-VM_LOCKED VMAs.
1433 *
1434 * In most cases the pmd is the only mapping of the page as we
1435 * break COW for the mlock() -- see gup_flags |= FOLL_WRITE for
1436 * writable private mappings in populate_vma_page_range().
1437 *
1438 * The only scenario when we have the page shared here is if we
1439 * mlocking read-only mapping shared over fork(). We skip
1440 * mlocking such pages.
1441 */
1442 if (compound_mapcount(page) == 1 && !PageDoubleMap(page) &&
1443 page->mapping && trylock_page(page)) {
1444 lru_add_drain();
1445 if (page->mapping)
1446 mlock_vma_page(page);
1447 unlock_page(page);
1448 }
1449 }
1450 page += (addr & ~HPAGE_PMD_MASK) >> PAGE_SHIFT;
1451 VM_BUG_ON_PAGE(!PageCompound(page), page);
1452 if (flags & FOLL_GET)
1453 get_page(page);
1454
1455out:
1456 return page;
1457}
1458
1459/* NUMA hinting page fault entry point for trans huge pmds */
1460int do_huge_pmd_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
1461 unsigned long addr, pmd_t pmd, pmd_t *pmdp)
1462{
1463 spinlock_t *ptl;
1464 struct anon_vma *anon_vma = NULL;
1465 struct page *page;
1466 unsigned long haddr = addr & HPAGE_PMD_MASK;
1467 int page_nid = -1, this_nid = numa_node_id();
1468 int target_nid, last_cpupid = -1;
1469 bool page_locked;
1470 bool migrated = false;
1471 bool was_writable;
1472 int flags = 0;
1473
1474 /* A PROT_NONE fault should not end up here */
1475 BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
1476
1477 ptl = pmd_lock(mm, pmdp);
1478 if (unlikely(!pmd_same(pmd, *pmdp)))
1479 goto out_unlock;
1480
1481 /*
1482 * If there are potential migrations, wait for completion and retry
1483 * without disrupting NUMA hinting information. Do not relock and
1484 * check_same as the page may no longer be mapped.
1485 */
1486 if (unlikely(pmd_trans_migrating(*pmdp))) {
1487 page = pmd_page(*pmdp);
1488 spin_unlock(ptl);
1489 wait_on_page_locked(page);
1490 goto out;
1491 }
1492
1493 page = pmd_page(pmd);
1494 BUG_ON(is_huge_zero_page(page));
1495 page_nid = page_to_nid(page);
1496 last_cpupid = page_cpupid_last(page);
1497 count_vm_numa_event(NUMA_HINT_FAULTS);
1498 if (page_nid == this_nid) {
1499 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
1500 flags |= TNF_FAULT_LOCAL;
1501 }
1502
1503 /* See similar comment in do_numa_page for explanation */
1504 if (!(vma->vm_flags & VM_WRITE))
1505 flags |= TNF_NO_GROUP;
1506
1507 /*
1508 * Acquire the page lock to serialise THP migrations but avoid dropping
1509 * page_table_lock if at all possible
1510 */
1511 page_locked = trylock_page(page);
1512 target_nid = mpol_misplaced(page, vma, haddr);
1513 if (target_nid == -1) {
1514 /* If the page was locked, there are no parallel migrations */
1515 if (page_locked)
1516 goto clear_pmdnuma;
1517 }
1518
1519 /* Migration could have started since the pmd_trans_migrating check */
1520 if (!page_locked) {
1521 spin_unlock(ptl);
1522 wait_on_page_locked(page);
1523 page_nid = -1;
1524 goto out;
1525 }
1526
1527 /*
1528 * Page is misplaced. Page lock serialises migrations. Acquire anon_vma
1529 * to serialises splits
1530 */
1531 get_page(page);
1532 spin_unlock(ptl);
1533 anon_vma = page_lock_anon_vma_read(page);
1534
1535 /* Confirm the PMD did not change while page_table_lock was released */
1536 spin_lock(ptl);
1537 if (unlikely(!pmd_same(pmd, *pmdp))) {
1538 unlock_page(page);
1539 put_page(page);
1540 page_nid = -1;
1541 goto out_unlock;
1542 }
1543
1544 /* Bail if we fail to protect against THP splits for any reason */
1545 if (unlikely(!anon_vma)) {
1546 put_page(page);
1547 page_nid = -1;
1548 goto clear_pmdnuma;
1549 }
1550
1551 /*
1552 * Migrate the THP to the requested node, returns with page unlocked
1553 * and access rights restored.
1554 */
1555 spin_unlock(ptl);
1556 migrated = migrate_misplaced_transhuge_page(mm, vma,
1557 pmdp, pmd, addr, page, target_nid);
1558 if (migrated) {
1559 flags |= TNF_MIGRATED;
1560 page_nid = target_nid;
1561 } else
1562 flags |= TNF_MIGRATE_FAIL;
1563
1564 goto out;
1565clear_pmdnuma:
1566 BUG_ON(!PageLocked(page));
1567 was_writable = pmd_write(pmd);
1568 pmd = pmd_modify(pmd, vma->vm_page_prot);
1569 pmd = pmd_mkyoung(pmd);
1570 if (was_writable)
1571 pmd = pmd_mkwrite(pmd);
1572 set_pmd_at(mm, haddr, pmdp, pmd);
1573 update_mmu_cache_pmd(vma, addr, pmdp);
1574 unlock_page(page);
1575out_unlock:
1576 spin_unlock(ptl);
1577
1578out:
1579 if (anon_vma)
1580 page_unlock_anon_vma_read(anon_vma);
1581
1582 if (page_nid != -1)
1583 task_numa_fault(last_cpupid, page_nid, HPAGE_PMD_NR, flags);
1584
1585 return 0;
1586}
1587
1588int madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1589 pmd_t *pmd, unsigned long addr, unsigned long next)
1590
1591{
1592 spinlock_t *ptl;
1593 pmd_t orig_pmd;
1594 struct page *page;
1595 struct mm_struct *mm = tlb->mm;
1596 int ret = 0;
1597
1598 ptl = pmd_trans_huge_lock(pmd, vma);
1599 if (!ptl)
1600 goto out_unlocked;
1601
1602 orig_pmd = *pmd;
1603 if (is_huge_zero_pmd(orig_pmd)) {
1604 ret = 1;
1605 goto out;
1606 }
1607
1608 page = pmd_page(orig_pmd);
1609 /*
1610 * If other processes are mapping this page, we couldn't discard
1611 * the page unless they all do MADV_FREE so let's skip the page.
1612 */
1613 if (page_mapcount(page) != 1)
1614 goto out;
1615
1616 if (!trylock_page(page))
1617 goto out;
1618
1619 /*
1620 * If user want to discard part-pages of THP, split it so MADV_FREE
1621 * will deactivate only them.
1622 */
1623 if (next - addr != HPAGE_PMD_SIZE) {
1624 get_page(page);
1625 spin_unlock(ptl);
1626 if (split_huge_page(page)) {
1627 put_page(page);
1628 unlock_page(page);
1629 goto out_unlocked;
1630 }
1631 put_page(page);
1632 unlock_page(page);
1633 ret = 1;
1634 goto out_unlocked;
1635 }
1636
1637 if (PageDirty(page))
1638 ClearPageDirty(page);
1639 unlock_page(page);
1640
1641 if (PageActive(page))
1642 deactivate_page(page);
1643
1644 if (pmd_young(orig_pmd) || pmd_dirty(orig_pmd)) {
1645 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1646 tlb->fullmm);
1647 orig_pmd = pmd_mkold(orig_pmd);
1648 orig_pmd = pmd_mkclean(orig_pmd);
1649
1650 set_pmd_at(mm, addr, pmd, orig_pmd);
1651 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1652 }
1653 ret = 1;
1654out:
1655 spin_unlock(ptl);
1656out_unlocked:
1657 return ret;
1658}
1659
1660int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
1661 pmd_t *pmd, unsigned long addr)
1662{
1663 pmd_t orig_pmd;
1664 spinlock_t *ptl;
1665
1666 ptl = __pmd_trans_huge_lock(pmd, vma);
1667 if (!ptl)
1668 return 0;
1669 /*
1670 * For architectures like ppc64 we look at deposited pgtable
1671 * when calling pmdp_huge_get_and_clear. So do the
1672 * pgtable_trans_huge_withdraw after finishing pmdp related
1673 * operations.
1674 */
1675 orig_pmd = pmdp_huge_get_and_clear_full(tlb->mm, addr, pmd,
1676 tlb->fullmm);
1677 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
1678 if (vma_is_dax(vma)) {
1679 spin_unlock(ptl);
1680 if (is_huge_zero_pmd(orig_pmd))
1681 tlb_remove_page(tlb, pmd_page(orig_pmd));
1682 } else if (is_huge_zero_pmd(orig_pmd)) {
1683 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1684 atomic_long_dec(&tlb->mm->nr_ptes);
1685 spin_unlock(ptl);
1686 tlb_remove_page(tlb, pmd_page(orig_pmd));
1687 } else {
1688 struct page *page = pmd_page(orig_pmd);
1689 page_remove_rmap(page, true);
1690 VM_BUG_ON_PAGE(page_mapcount(page) < 0, page);
1691 add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
1692 VM_BUG_ON_PAGE(!PageHead(page), page);
1693 pte_free(tlb->mm, pgtable_trans_huge_withdraw(tlb->mm, pmd));
1694 atomic_long_dec(&tlb->mm->nr_ptes);
1695 spin_unlock(ptl);
1696 tlb_remove_page(tlb, page);
1697 }
1698 return 1;
1699}
1700
1701bool move_huge_pmd(struct vm_area_struct *vma, struct vm_area_struct *new_vma,
1702 unsigned long old_addr,
1703 unsigned long new_addr, unsigned long old_end,
1704 pmd_t *old_pmd, pmd_t *new_pmd)
1705{
1706 spinlock_t *old_ptl, *new_ptl;
1707 pmd_t pmd;
1708
1709 struct mm_struct *mm = vma->vm_mm;
1710
1711 if ((old_addr & ~HPAGE_PMD_MASK) ||
1712 (new_addr & ~HPAGE_PMD_MASK) ||
1713 old_end - old_addr < HPAGE_PMD_SIZE ||
1714 (new_vma->vm_flags & VM_NOHUGEPAGE))
1715 return false;
1716
1717 /*
1718 * The destination pmd shouldn't be established, free_pgtables()
1719 * should have release it.
1720 */
1721 if (WARN_ON(!pmd_none(*new_pmd))) {
1722 VM_BUG_ON(pmd_trans_huge(*new_pmd));
1723 return false;
1724 }
1725
1726 /*
1727 * We don't have to worry about the ordering of src and dst
1728 * ptlocks because exclusive mmap_sem prevents deadlock.
1729 */
1730 old_ptl = __pmd_trans_huge_lock(old_pmd, vma);
1731 if (old_ptl) {
1732 new_ptl = pmd_lockptr(mm, new_pmd);
1733 if (new_ptl != old_ptl)
1734 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
1735 pmd = pmdp_huge_get_and_clear(mm, old_addr, old_pmd);
1736 VM_BUG_ON(!pmd_none(*new_pmd));
1737
1738 if (pmd_move_must_withdraw(new_ptl, old_ptl) &&
1739 vma_is_anonymous(vma)) {
1740 pgtable_t pgtable;
1741 pgtable = pgtable_trans_huge_withdraw(mm, old_pmd);
1742 pgtable_trans_huge_deposit(mm, new_pmd, pgtable);
1743 }
1744 set_pmd_at(mm, new_addr, new_pmd, pmd_mksoft_dirty(pmd));
1745 if (new_ptl != old_ptl)
1746 spin_unlock(new_ptl);
1747 spin_unlock(old_ptl);
1748 return true;
1749 }
1750 return false;
1751}
1752
1753/*
1754 * Returns
1755 * - 0 if PMD could not be locked
1756 * - 1 if PMD was locked but protections unchange and TLB flush unnecessary
1757 * - HPAGE_PMD_NR is protections changed and TLB flush necessary
1758 */
1759int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
1760 unsigned long addr, pgprot_t newprot, int prot_numa)
1761{
1762 struct mm_struct *mm = vma->vm_mm;
1763 spinlock_t *ptl;
1764 int ret = 0;
1765
1766 ptl = __pmd_trans_huge_lock(pmd, vma);
1767 if (ptl) {
1768 pmd_t entry;
1769 bool preserve_write = prot_numa && pmd_write(*pmd);
1770 ret = 1;
1771
1772 /*
1773 * Avoid trapping faults against the zero page. The read-only
1774 * data is likely to be read-cached on the local CPU and
1775 * local/remote hits to the zero page are not interesting.
1776 */
1777 if (prot_numa && is_huge_zero_pmd(*pmd)) {
1778 spin_unlock(ptl);
1779 return ret;
1780 }
1781
1782 if (!prot_numa || !pmd_protnone(*pmd)) {
1783 entry = pmdp_huge_get_and_clear_notify(mm, addr, pmd);
1784 entry = pmd_modify(entry, newprot);
1785 if (preserve_write)
1786 entry = pmd_mkwrite(entry);
1787 ret = HPAGE_PMD_NR;
1788 set_pmd_at(mm, addr, pmd, entry);
1789 BUG_ON(!preserve_write && pmd_write(entry));
1790 }
1791 spin_unlock(ptl);
1792 }
1793
1794 return ret;
1795}
1796
1797/*
1798 * Returns true if a given pmd maps a thp, false otherwise.
1799 *
1800 * Note that if it returns true, this routine returns without unlocking page
1801 * table lock. So callers must unlock it.
1802 */
1803spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma)
1804{
1805 spinlock_t *ptl;
1806 ptl = pmd_lock(vma->vm_mm, pmd);
1807 if (likely(pmd_trans_huge(*pmd) || pmd_devmap(*pmd)))
1808 return ptl;
1809 spin_unlock(ptl);
1810 return NULL;
1811}
1812
1813#define VM_NO_THP (VM_SPECIAL | VM_HUGETLB | VM_SHARED | VM_MAYSHARE)
1814
1815int hugepage_madvise(struct vm_area_struct *vma,
1816 unsigned long *vm_flags, int advice)
1817{
1818 switch (advice) {
1819 case MADV_HUGEPAGE:
1820#ifdef CONFIG_S390
1821 /*
1822 * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
1823 * can't handle this properly after s390_enable_sie, so we simply
1824 * ignore the madvise to prevent qemu from causing a SIGSEGV.
1825 */
1826 if (mm_has_pgste(vma->vm_mm))
1827 return 0;
1828#endif
1829 /*
1830 * Be somewhat over-protective like KSM for now!
1831 */
1832 if (*vm_flags & VM_NO_THP)
1833 return -EINVAL;
1834 *vm_flags &= ~VM_NOHUGEPAGE;
1835 *vm_flags |= VM_HUGEPAGE;
1836 /*
1837 * If the vma become good for khugepaged to scan,
1838 * register it here without waiting a page fault that
1839 * may not happen any time soon.
1840 */
1841 if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
1842 return -ENOMEM;
1843 break;
1844 case MADV_NOHUGEPAGE:
1845 /*
1846 * Be somewhat over-protective like KSM for now!
1847 */
1848 if (*vm_flags & VM_NO_THP)
1849 return -EINVAL;
1850 *vm_flags &= ~VM_HUGEPAGE;
1851 *vm_flags |= VM_NOHUGEPAGE;
1852 /*
1853 * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
1854 * this vma even if we leave the mm registered in khugepaged if
1855 * it got registered before VM_NOHUGEPAGE was set.
1856 */
1857 break;
1858 }
1859
1860 return 0;
1861}
1862
1863static int __init khugepaged_slab_init(void)
1864{
1865 mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
1866 sizeof(struct mm_slot),
1867 __alignof__(struct mm_slot), 0, NULL);
1868 if (!mm_slot_cache)
1869 return -ENOMEM;
1870
1871 return 0;
1872}
1873
1874static void __init khugepaged_slab_exit(void)
1875{
1876 kmem_cache_destroy(mm_slot_cache);
1877}
1878
1879static inline struct mm_slot *alloc_mm_slot(void)
1880{
1881 if (!mm_slot_cache) /* initialization failed */
1882 return NULL;
1883 return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
1884}
1885
1886static inline void free_mm_slot(struct mm_slot *mm_slot)
1887{
1888 kmem_cache_free(mm_slot_cache, mm_slot);
1889}
1890
1891static struct mm_slot *get_mm_slot(struct mm_struct *mm)
1892{
1893 struct mm_slot *mm_slot;
1894
1895 hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
1896 if (mm == mm_slot->mm)
1897 return mm_slot;
1898
1899 return NULL;
1900}
1901
1902static void insert_to_mm_slots_hash(struct mm_struct *mm,
1903 struct mm_slot *mm_slot)
1904{
1905 mm_slot->mm = mm;
1906 hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
1907}
1908
1909static inline int khugepaged_test_exit(struct mm_struct *mm)
1910{
1911 return atomic_read(&mm->mm_users) == 0;
1912}
1913
1914int __khugepaged_enter(struct mm_struct *mm)
1915{
1916 struct mm_slot *mm_slot;
1917 int wakeup;
1918
1919 mm_slot = alloc_mm_slot();
1920 if (!mm_slot)
1921 return -ENOMEM;
1922
1923 /* __khugepaged_exit() must not run from under us */
1924 VM_BUG_ON_MM(khugepaged_test_exit(mm), mm);
1925 if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
1926 free_mm_slot(mm_slot);
1927 return 0;
1928 }
1929
1930 spin_lock(&khugepaged_mm_lock);
1931 insert_to_mm_slots_hash(mm, mm_slot);
1932 /*
1933 * Insert just behind the scanning cursor, to let the area settle
1934 * down a little.
1935 */
1936 wakeup = list_empty(&khugepaged_scan.mm_head);
1937 list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
1938 spin_unlock(&khugepaged_mm_lock);
1939
1940 atomic_inc(&mm->mm_count);
1941 if (wakeup)
1942 wake_up_interruptible(&khugepaged_wait);
1943
1944 return 0;
1945}
1946
1947int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
1948 unsigned long vm_flags)
1949{
1950 unsigned long hstart, hend;
1951 if (!vma->anon_vma)
1952 /*
1953 * Not yet faulted in so we will register later in the
1954 * page fault if needed.
1955 */
1956 return 0;
1957 if (vma->vm_ops || (vm_flags & VM_NO_THP))
1958 /* khugepaged not yet working on file or special mappings */
1959 return 0;
1960 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1961 hend = vma->vm_end & HPAGE_PMD_MASK;
1962 if (hstart < hend)
1963 return khugepaged_enter(vma, vm_flags);
1964 return 0;
1965}
1966
1967void __khugepaged_exit(struct mm_struct *mm)
1968{
1969 struct mm_slot *mm_slot;
1970 int free = 0;
1971
1972 spin_lock(&khugepaged_mm_lock);
1973 mm_slot = get_mm_slot(mm);
1974 if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
1975 hash_del(&mm_slot->hash);
1976 list_del(&mm_slot->mm_node);
1977 free = 1;
1978 }
1979 spin_unlock(&khugepaged_mm_lock);
1980
1981 if (free) {
1982 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1983 free_mm_slot(mm_slot);
1984 mmdrop(mm);
1985 } else if (mm_slot) {
1986 /*
1987 * This is required to serialize against
1988 * khugepaged_test_exit() (which is guaranteed to run
1989 * under mmap sem read mode). Stop here (after we
1990 * return all pagetables will be destroyed) until
1991 * khugepaged has finished working on the pagetables
1992 * under the mmap_sem.
1993 */
1994 down_write(&mm->mmap_sem);
1995 up_write(&mm->mmap_sem);
1996 }
1997}
1998
1999static void release_pte_page(struct page *page)
2000{
2001 /* 0 stands for page_is_file_cache(page) == false */
2002 dec_zone_page_state(page, NR_ISOLATED_ANON + 0);
2003 unlock_page(page);
2004 putback_lru_page(page);
2005}
2006
2007static void release_pte_pages(pte_t *pte, pte_t *_pte)
2008{
2009 while (--_pte >= pte) {
2010 pte_t pteval = *_pte;
2011 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
2012 release_pte_page(pte_page(pteval));
2013 }
2014}
2015
2016static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
2017 unsigned long address,
2018 pte_t *pte)
2019{
2020 struct page *page = NULL;
2021 pte_t *_pte;
2022 int none_or_zero = 0, result = 0;
2023 bool referenced = false, writable = false;
2024
2025 for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
2026 _pte++, address += PAGE_SIZE) {
2027 pte_t pteval = *_pte;
2028 if (pte_none(pteval) || (pte_present(pteval) &&
2029 is_zero_pfn(pte_pfn(pteval)))) {
2030 if (!userfaultfd_armed(vma) &&
2031 ++none_or_zero <= khugepaged_max_ptes_none) {
2032 continue;
2033 } else {
2034 result = SCAN_EXCEED_NONE_PTE;
2035 goto out;
2036 }
2037 }
2038 if (!pte_present(pteval)) {
2039 result = SCAN_PTE_NON_PRESENT;
2040 goto out;
2041 }
2042 page = vm_normal_page(vma, address, pteval);
2043 if (unlikely(!page)) {
2044 result = SCAN_PAGE_NULL;
2045 goto out;
2046 }
2047
2048 VM_BUG_ON_PAGE(PageCompound(page), page);
2049 VM_BUG_ON_PAGE(!PageAnon(page), page);
2050 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
2051
2052 /*
2053 * We can do it before isolate_lru_page because the
2054 * page can't be freed from under us. NOTE: PG_lock
2055 * is needed to serialize against split_huge_page
2056 * when invoked from the VM.
2057 */
2058 if (!trylock_page(page)) {
2059 result = SCAN_PAGE_LOCK;
2060 goto out;
2061 }
2062
2063 /*
2064 * cannot use mapcount: can't collapse if there's a gup pin.
2065 * The page must only be referenced by the scanned process
2066 * and page swap cache.
2067 */
2068 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2069 unlock_page(page);
2070 result = SCAN_PAGE_COUNT;
2071 goto out;
2072 }
2073 if (pte_write(pteval)) {
2074 writable = true;
2075 } else {
2076 if (PageSwapCache(page) &&
2077 !reuse_swap_page(page, NULL)) {
2078 unlock_page(page);
2079 result = SCAN_SWAP_CACHE_PAGE;
2080 goto out;
2081 }
2082 /*
2083 * Page is not in the swap cache. It can be collapsed
2084 * into a THP.
2085 */
2086 }
2087
2088 /*
2089 * Isolate the page to avoid collapsing an hugepage
2090 * currently in use by the VM.
2091 */
2092 if (isolate_lru_page(page)) {
2093 unlock_page(page);
2094 result = SCAN_DEL_PAGE_LRU;
2095 goto out;
2096 }
2097 /* 0 stands for page_is_file_cache(page) == false */
2098 inc_zone_page_state(page, NR_ISOLATED_ANON + 0);
2099 VM_BUG_ON_PAGE(!PageLocked(page), page);
2100 VM_BUG_ON_PAGE(PageLRU(page), page);
2101
2102 /* If there is no mapped pte young don't collapse the page */
2103 if (pte_young(pteval) ||
2104 page_is_young(page) || PageReferenced(page) ||
2105 mmu_notifier_test_young(vma->vm_mm, address))
2106 referenced = true;
2107 }
2108 if (likely(writable)) {
2109 if (likely(referenced)) {
2110 result = SCAN_SUCCEED;
2111 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
2112 referenced, writable, result);
2113 return 1;
2114 }
2115 } else {
2116 result = SCAN_PAGE_RO;
2117 }
2118
2119out:
2120 release_pte_pages(pte, _pte);
2121 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
2122 referenced, writable, result);
2123 return 0;
2124}
2125
2126static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
2127 struct vm_area_struct *vma,
2128 unsigned long address,
2129 spinlock_t *ptl)
2130{
2131 pte_t *_pte;
2132 for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
2133 pte_t pteval = *_pte;
2134 struct page *src_page;
2135
2136 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2137 clear_user_highpage(page, address);
2138 add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
2139 if (is_zero_pfn(pte_pfn(pteval))) {
2140 /*
2141 * ptl mostly unnecessary.
2142 */
2143 spin_lock(ptl);
2144 /*
2145 * paravirt calls inside pte_clear here are
2146 * superfluous.
2147 */
2148 pte_clear(vma->vm_mm, address, _pte);
2149 spin_unlock(ptl);
2150 }
2151 } else {
2152 src_page = pte_page(pteval);
2153 copy_user_highpage(page, src_page, address, vma);
2154 VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
2155 release_pte_page(src_page);
2156 /*
2157 * ptl mostly unnecessary, but preempt has to
2158 * be disabled to update the per-cpu stats
2159 * inside page_remove_rmap().
2160 */
2161 spin_lock(ptl);
2162 /*
2163 * paravirt calls inside pte_clear here are
2164 * superfluous.
2165 */
2166 pte_clear(vma->vm_mm, address, _pte);
2167 page_remove_rmap(src_page, false);
2168 spin_unlock(ptl);
2169 free_page_and_swap_cache(src_page);
2170 }
2171
2172 address += PAGE_SIZE;
2173 page++;
2174 }
2175}
2176
2177static void khugepaged_alloc_sleep(void)
2178{
2179 DEFINE_WAIT(wait);
2180
2181 add_wait_queue(&khugepaged_wait, &wait);
2182 freezable_schedule_timeout_interruptible(
2183 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
2184 remove_wait_queue(&khugepaged_wait, &wait);
2185}
2186
2187static int khugepaged_node_load[MAX_NUMNODES];
2188
2189static bool khugepaged_scan_abort(int nid)
2190{
2191 int i;
2192
2193 /*
2194 * If zone_reclaim_mode is disabled, then no extra effort is made to
2195 * allocate memory locally.
2196 */
2197 if (!zone_reclaim_mode)
2198 return false;
2199
2200 /* If there is a count for this node already, it must be acceptable */
2201 if (khugepaged_node_load[nid])
2202 return false;
2203
2204 for (i = 0; i < MAX_NUMNODES; i++) {
2205 if (!khugepaged_node_load[i])
2206 continue;
2207 if (node_distance(nid, i) > RECLAIM_DISTANCE)
2208 return true;
2209 }
2210 return false;
2211}
2212
2213#ifdef CONFIG_NUMA
2214static int khugepaged_find_target_node(void)
2215{
2216 static int last_khugepaged_target_node = NUMA_NO_NODE;
2217 int nid, target_node = 0, max_value = 0;
2218
2219 /* find first node with max normal pages hit */
2220 for (nid = 0; nid < MAX_NUMNODES; nid++)
2221 if (khugepaged_node_load[nid] > max_value) {
2222 max_value = khugepaged_node_load[nid];
2223 target_node = nid;
2224 }
2225
2226 /* do some balance if several nodes have the same hit record */
2227 if (target_node <= last_khugepaged_target_node)
2228 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
2229 nid++)
2230 if (max_value == khugepaged_node_load[nid]) {
2231 target_node = nid;
2232 break;
2233 }
2234
2235 last_khugepaged_target_node = target_node;
2236 return target_node;
2237}
2238
2239static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2240{
2241 if (IS_ERR(*hpage)) {
2242 if (!*wait)
2243 return false;
2244
2245 *wait = false;
2246 *hpage = NULL;
2247 khugepaged_alloc_sleep();
2248 } else if (*hpage) {
2249 put_page(*hpage);
2250 *hpage = NULL;
2251 }
2252
2253 return true;
2254}
2255
2256static struct page *
2257khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2258 unsigned long address, int node)
2259{
2260 VM_BUG_ON_PAGE(*hpage, *hpage);
2261
2262 /*
2263 * Before allocating the hugepage, release the mmap_sem read lock.
2264 * The allocation can take potentially a long time if it involves
2265 * sync compaction, and we do not need to hold the mmap_sem during
2266 * that. We will recheck the vma after taking it again in write mode.
2267 */
2268 up_read(&mm->mmap_sem);
2269
2270 *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
2271 if (unlikely(!*hpage)) {
2272 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2273 *hpage = ERR_PTR(-ENOMEM);
2274 return NULL;
2275 }
2276
2277 prep_transhuge_page(*hpage);
2278 count_vm_event(THP_COLLAPSE_ALLOC);
2279 return *hpage;
2280}
2281#else
2282static int khugepaged_find_target_node(void)
2283{
2284 return 0;
2285}
2286
2287static inline struct page *alloc_khugepaged_hugepage(void)
2288{
2289 struct page *page;
2290
2291 page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
2292 HPAGE_PMD_ORDER);
2293 if (page)
2294 prep_transhuge_page(page);
2295 return page;
2296}
2297
2298static struct page *khugepaged_alloc_hugepage(bool *wait)
2299{
2300 struct page *hpage;
2301
2302 do {
2303 hpage = alloc_khugepaged_hugepage();
2304 if (!hpage) {
2305 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
2306 if (!*wait)
2307 return NULL;
2308
2309 *wait = false;
2310 khugepaged_alloc_sleep();
2311 } else
2312 count_vm_event(THP_COLLAPSE_ALLOC);
2313 } while (unlikely(!hpage) && likely(khugepaged_enabled()));
2314
2315 return hpage;
2316}
2317
2318static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
2319{
2320 if (!*hpage)
2321 *hpage = khugepaged_alloc_hugepage(wait);
2322
2323 if (unlikely(!*hpage))
2324 return false;
2325
2326 return true;
2327}
2328
2329static struct page *
2330khugepaged_alloc_page(struct page **hpage, gfp_t gfp, struct mm_struct *mm,
2331 unsigned long address, int node)
2332{
2333 up_read(&mm->mmap_sem);
2334 VM_BUG_ON(!*hpage);
2335
2336 return *hpage;
2337}
2338#endif
2339
2340static bool hugepage_vma_check(struct vm_area_struct *vma)
2341{
2342 if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
2343 (vma->vm_flags & VM_NOHUGEPAGE))
2344 return false;
2345 if (!vma->anon_vma || vma->vm_ops)
2346 return false;
2347 if (is_vma_temporary_stack(vma))
2348 return false;
2349 return !(vma->vm_flags & VM_NO_THP);
2350}
2351
2352static void collapse_huge_page(struct mm_struct *mm,
2353 unsigned long address,
2354 struct page **hpage,
2355 struct vm_area_struct *vma,
2356 int node)
2357{
2358 pmd_t *pmd, _pmd;
2359 pte_t *pte;
2360 pgtable_t pgtable;
2361 struct page *new_page;
2362 spinlock_t *pmd_ptl, *pte_ptl;
2363 int isolated = 0, result = 0;
2364 unsigned long hstart, hend;
2365 struct mem_cgroup *memcg;
2366 unsigned long mmun_start; /* For mmu_notifiers */
2367 unsigned long mmun_end; /* For mmu_notifiers */
2368 gfp_t gfp;
2369
2370 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2371
2372 /* Only allocate from the target node */
2373 gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_OTHER_NODE | __GFP_THISNODE;
2374
2375 /* release the mmap_sem read lock. */
2376 new_page = khugepaged_alloc_page(hpage, gfp, mm, address, node);
2377 if (!new_page) {
2378 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
2379 goto out_nolock;
2380 }
2381
2382 if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp, &memcg, true))) {
2383 result = SCAN_CGROUP_CHARGE_FAIL;
2384 goto out_nolock;
2385 }
2386
2387 /*
2388 * Prevent all access to pagetables with the exception of
2389 * gup_fast later hanlded by the ptep_clear_flush and the VM
2390 * handled by the anon_vma lock + PG_lock.
2391 */
2392 down_write(&mm->mmap_sem);
2393 if (unlikely(khugepaged_test_exit(mm))) {
2394 result = SCAN_ANY_PROCESS;
2395 goto out;
2396 }
2397
2398 vma = find_vma(mm, address);
2399 if (!vma) {
2400 result = SCAN_VMA_NULL;
2401 goto out;
2402 }
2403 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2404 hend = vma->vm_end & HPAGE_PMD_MASK;
2405 if (address < hstart || address + HPAGE_PMD_SIZE > hend) {
2406 result = SCAN_ADDRESS_RANGE;
2407 goto out;
2408 }
2409 if (!hugepage_vma_check(vma)) {
2410 result = SCAN_VMA_CHECK;
2411 goto out;
2412 }
2413 pmd = mm_find_pmd(mm, address);
2414 if (!pmd) {
2415 result = SCAN_PMD_NULL;
2416 goto out;
2417 }
2418
2419 anon_vma_lock_write(vma->anon_vma);
2420
2421 pte = pte_offset_map(pmd, address);
2422 pte_ptl = pte_lockptr(mm, pmd);
2423
2424 mmun_start = address;
2425 mmun_end = address + HPAGE_PMD_SIZE;
2426 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2427 pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
2428 /*
2429 * After this gup_fast can't run anymore. This also removes
2430 * any huge TLB entry from the CPU so we won't allow
2431 * huge and small TLB entries for the same virtual address
2432 * to avoid the risk of CPU bugs in that area.
2433 */
2434 _pmd = pmdp_collapse_flush(vma, address, pmd);
2435 spin_unlock(pmd_ptl);
2436 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2437
2438 spin_lock(pte_ptl);
2439 isolated = __collapse_huge_page_isolate(vma, address, pte);
2440 spin_unlock(pte_ptl);
2441
2442 if (unlikely(!isolated)) {
2443 pte_unmap(pte);
2444 spin_lock(pmd_ptl);
2445 BUG_ON(!pmd_none(*pmd));
2446 /*
2447 * We can only use set_pmd_at when establishing
2448 * hugepmds and never for establishing regular pmds that
2449 * points to regular pagetables. Use pmd_populate for that
2450 */
2451 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
2452 spin_unlock(pmd_ptl);
2453 anon_vma_unlock_write(vma->anon_vma);
2454 result = SCAN_FAIL;
2455 goto out;
2456 }
2457
2458 /*
2459 * All pages are isolated and locked so anon_vma rmap
2460 * can't run anymore.
2461 */
2462 anon_vma_unlock_write(vma->anon_vma);
2463
2464 __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
2465 pte_unmap(pte);
2466 __SetPageUptodate(new_page);
2467 pgtable = pmd_pgtable(_pmd);
2468
2469 _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
2470 _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
2471
2472 /*
2473 * spin_lock() below is not the equivalent of smp_wmb(), so
2474 * this is needed to avoid the copy_huge_page writes to become
2475 * visible after the set_pmd_at() write.
2476 */
2477 smp_wmb();
2478
2479 spin_lock(pmd_ptl);
2480 BUG_ON(!pmd_none(*pmd));
2481 page_add_new_anon_rmap(new_page, vma, address, true);
2482 mem_cgroup_commit_charge(new_page, memcg, false, true);
2483 lru_cache_add_active_or_unevictable(new_page, vma);
2484 pgtable_trans_huge_deposit(mm, pmd, pgtable);
2485 set_pmd_at(mm, address, pmd, _pmd);
2486 update_mmu_cache_pmd(vma, address, pmd);
2487 spin_unlock(pmd_ptl);
2488
2489 *hpage = NULL;
2490
2491 khugepaged_pages_collapsed++;
2492 result = SCAN_SUCCEED;
2493out_up_write:
2494 up_write(&mm->mmap_sem);
2495 trace_mm_collapse_huge_page(mm, isolated, result);
2496 return;
2497
2498out_nolock:
2499 trace_mm_collapse_huge_page(mm, isolated, result);
2500 return;
2501out:
2502 mem_cgroup_cancel_charge(new_page, memcg, true);
2503 goto out_up_write;
2504}
2505
2506static int khugepaged_scan_pmd(struct mm_struct *mm,
2507 struct vm_area_struct *vma,
2508 unsigned long address,
2509 struct page **hpage)
2510{
2511 pmd_t *pmd;
2512 pte_t *pte, *_pte;
2513 int ret = 0, none_or_zero = 0, result = 0;
2514 struct page *page = NULL;
2515 unsigned long _address;
2516 spinlock_t *ptl;
2517 int node = NUMA_NO_NODE;
2518 bool writable = false, referenced = false;
2519
2520 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
2521
2522 pmd = mm_find_pmd(mm, address);
2523 if (!pmd) {
2524 result = SCAN_PMD_NULL;
2525 goto out;
2526 }
2527
2528 memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
2529 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2530 for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
2531 _pte++, _address += PAGE_SIZE) {
2532 pte_t pteval = *_pte;
2533 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
2534 if (!userfaultfd_armed(vma) &&
2535 ++none_or_zero <= khugepaged_max_ptes_none) {
2536 continue;
2537 } else {
2538 result = SCAN_EXCEED_NONE_PTE;
2539 goto out_unmap;
2540 }
2541 }
2542 if (!pte_present(pteval)) {
2543 result = SCAN_PTE_NON_PRESENT;
2544 goto out_unmap;
2545 }
2546 if (pte_write(pteval))
2547 writable = true;
2548
2549 page = vm_normal_page(vma, _address, pteval);
2550 if (unlikely(!page)) {
2551 result = SCAN_PAGE_NULL;
2552 goto out_unmap;
2553 }
2554
2555 /* TODO: teach khugepaged to collapse THP mapped with pte */
2556 if (PageCompound(page)) {
2557 result = SCAN_PAGE_COMPOUND;
2558 goto out_unmap;
2559 }
2560
2561 /*
2562 * Record which node the original page is from and save this
2563 * information to khugepaged_node_load[].
2564 * Khupaged will allocate hugepage from the node has the max
2565 * hit record.
2566 */
2567 node = page_to_nid(page);
2568 if (khugepaged_scan_abort(node)) {
2569 result = SCAN_SCAN_ABORT;
2570 goto out_unmap;
2571 }
2572 khugepaged_node_load[node]++;
2573 if (!PageLRU(page)) {
2574 result = SCAN_PAGE_LRU;
2575 goto out_unmap;
2576 }
2577 if (PageLocked(page)) {
2578 result = SCAN_PAGE_LOCK;
2579 goto out_unmap;
2580 }
2581 if (!PageAnon(page)) {
2582 result = SCAN_PAGE_ANON;
2583 goto out_unmap;
2584 }
2585
2586 /*
2587 * cannot use mapcount: can't collapse if there's a gup pin.
2588 * The page must only be referenced by the scanned process
2589 * and page swap cache.
2590 */
2591 if (page_count(page) != 1 + !!PageSwapCache(page)) {
2592 result = SCAN_PAGE_COUNT;
2593 goto out_unmap;
2594 }
2595 if (pte_young(pteval) ||
2596 page_is_young(page) || PageReferenced(page) ||
2597 mmu_notifier_test_young(vma->vm_mm, address))
2598 referenced = true;
2599 }
2600 if (writable) {
2601 if (referenced) {
2602 result = SCAN_SUCCEED;
2603 ret = 1;
2604 } else {
2605 result = SCAN_NO_REFERENCED_PAGE;
2606 }
2607 } else {
2608 result = SCAN_PAGE_RO;
2609 }
2610out_unmap:
2611 pte_unmap_unlock(pte, ptl);
2612 if (ret) {
2613 node = khugepaged_find_target_node();
2614 /* collapse_huge_page will return with the mmap_sem released */
2615 collapse_huge_page(mm, address, hpage, vma, node);
2616 }
2617out:
2618 trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
2619 none_or_zero, result);
2620 return ret;
2621}
2622
2623static void collect_mm_slot(struct mm_slot *mm_slot)
2624{
2625 struct mm_struct *mm = mm_slot->mm;
2626
2627 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2628
2629 if (khugepaged_test_exit(mm)) {
2630 /* free mm_slot */
2631 hash_del(&mm_slot->hash);
2632 list_del(&mm_slot->mm_node);
2633
2634 /*
2635 * Not strictly needed because the mm exited already.
2636 *
2637 * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
2638 */
2639
2640 /* khugepaged_mm_lock actually not necessary for the below */
2641 free_mm_slot(mm_slot);
2642 mmdrop(mm);
2643 }
2644}
2645
2646static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
2647 struct page **hpage)
2648 __releases(&khugepaged_mm_lock)
2649 __acquires(&khugepaged_mm_lock)
2650{
2651 struct mm_slot *mm_slot;
2652 struct mm_struct *mm;
2653 struct vm_area_struct *vma;
2654 int progress = 0;
2655
2656 VM_BUG_ON(!pages);
2657 VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
2658
2659 if (khugepaged_scan.mm_slot)
2660 mm_slot = khugepaged_scan.mm_slot;
2661 else {
2662 mm_slot = list_entry(khugepaged_scan.mm_head.next,
2663 struct mm_slot, mm_node);
2664 khugepaged_scan.address = 0;
2665 khugepaged_scan.mm_slot = mm_slot;
2666 }
2667 spin_unlock(&khugepaged_mm_lock);
2668
2669 mm = mm_slot->mm;
2670 down_read(&mm->mmap_sem);
2671 if (unlikely(khugepaged_test_exit(mm)))
2672 vma = NULL;
2673 else
2674 vma = find_vma(mm, khugepaged_scan.address);
2675
2676 progress++;
2677 for (; vma; vma = vma->vm_next) {
2678 unsigned long hstart, hend;
2679
2680 cond_resched();
2681 if (unlikely(khugepaged_test_exit(mm))) {
2682 progress++;
2683 break;
2684 }
2685 if (!hugepage_vma_check(vma)) {
2686skip:
2687 progress++;
2688 continue;
2689 }
2690 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2691 hend = vma->vm_end & HPAGE_PMD_MASK;
2692 if (hstart >= hend)
2693 goto skip;
2694 if (khugepaged_scan.address > hend)
2695 goto skip;
2696 if (khugepaged_scan.address < hstart)
2697 khugepaged_scan.address = hstart;
2698 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
2699
2700 while (khugepaged_scan.address < hend) {
2701 int ret;
2702 cond_resched();
2703 if (unlikely(khugepaged_test_exit(mm)))
2704 goto breakouterloop;
2705
2706 VM_BUG_ON(khugepaged_scan.address < hstart ||
2707 khugepaged_scan.address + HPAGE_PMD_SIZE >
2708 hend);
2709 ret = khugepaged_scan_pmd(mm, vma,
2710 khugepaged_scan.address,
2711 hpage);
2712 /* move to next address */
2713 khugepaged_scan.address += HPAGE_PMD_SIZE;
2714 progress += HPAGE_PMD_NR;
2715 if (ret)
2716 /* we released mmap_sem so break loop */
2717 goto breakouterloop_mmap_sem;
2718 if (progress >= pages)
2719 goto breakouterloop;
2720 }
2721 }
2722breakouterloop:
2723 up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
2724breakouterloop_mmap_sem:
2725
2726 spin_lock(&khugepaged_mm_lock);
2727 VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
2728 /*
2729 * Release the current mm_slot if this mm is about to die, or
2730 * if we scanned all vmas of this mm.
2731 */
2732 if (khugepaged_test_exit(mm) || !vma) {
2733 /*
2734 * Make sure that if mm_users is reaching zero while
2735 * khugepaged runs here, khugepaged_exit will find
2736 * mm_slot not pointing to the exiting mm.
2737 */
2738 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
2739 khugepaged_scan.mm_slot = list_entry(
2740 mm_slot->mm_node.next,
2741 struct mm_slot, mm_node);
2742 khugepaged_scan.address = 0;
2743 } else {
2744 khugepaged_scan.mm_slot = NULL;
2745 khugepaged_full_scans++;
2746 }
2747
2748 collect_mm_slot(mm_slot);
2749 }
2750
2751 return progress;
2752}
2753
2754static int khugepaged_has_work(void)
2755{
2756 return !list_empty(&khugepaged_scan.mm_head) &&
2757 khugepaged_enabled();
2758}
2759
2760static int khugepaged_wait_event(void)
2761{
2762 return !list_empty(&khugepaged_scan.mm_head) ||
2763 kthread_should_stop();
2764}
2765
2766static void khugepaged_do_scan(void)
2767{
2768 struct page *hpage = NULL;
2769 unsigned int progress = 0, pass_through_head = 0;
2770 unsigned int pages = khugepaged_pages_to_scan;
2771 bool wait = true;
2772
2773 barrier(); /* write khugepaged_pages_to_scan to local stack */
2774
2775 while (progress < pages) {
2776 if (!khugepaged_prealloc_page(&hpage, &wait))
2777 break;
2778
2779 cond_resched();
2780
2781 if (unlikely(kthread_should_stop() || try_to_freeze()))
2782 break;
2783
2784 spin_lock(&khugepaged_mm_lock);
2785 if (!khugepaged_scan.mm_slot)
2786 pass_through_head++;
2787 if (khugepaged_has_work() &&
2788 pass_through_head < 2)
2789 progress += khugepaged_scan_mm_slot(pages - progress,
2790 &hpage);
2791 else
2792 progress = pages;
2793 spin_unlock(&khugepaged_mm_lock);
2794 }
2795
2796 if (!IS_ERR_OR_NULL(hpage))
2797 put_page(hpage);
2798}
2799
2800static void khugepaged_wait_work(void)
2801{
2802 if (khugepaged_has_work()) {
2803 if (!khugepaged_scan_sleep_millisecs)
2804 return;
2805
2806 wait_event_freezable_timeout(khugepaged_wait,
2807 kthread_should_stop(),
2808 msecs_to_jiffies(khugepaged_scan_sleep_millisecs));
2809 return;
2810 }
2811
2812 if (khugepaged_enabled())
2813 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
2814}
2815
2816static int khugepaged(void *none)
2817{
2818 struct mm_slot *mm_slot;
2819
2820 set_freezable();
2821 set_user_nice(current, MAX_NICE);
2822
2823 while (!kthread_should_stop()) {
2824 khugepaged_do_scan();
2825 khugepaged_wait_work();
2826 }
2827
2828 spin_lock(&khugepaged_mm_lock);
2829 mm_slot = khugepaged_scan.mm_slot;
2830 khugepaged_scan.mm_slot = NULL;
2831 if (mm_slot)
2832 collect_mm_slot(mm_slot);
2833 spin_unlock(&khugepaged_mm_lock);
2834 return 0;
2835}
2836
2837static void __split_huge_zero_page_pmd(struct vm_area_struct *vma,
2838 unsigned long haddr, pmd_t *pmd)
2839{
2840 struct mm_struct *mm = vma->vm_mm;
2841 pgtable_t pgtable;
2842 pmd_t _pmd;
2843 int i;
2844
2845 /* leave pmd empty until pte is filled */
2846 pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2847
2848 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2849 pmd_populate(mm, &_pmd, pgtable);
2850
2851 for (i = 0; i < HPAGE_PMD_NR; i++, haddr += PAGE_SIZE) {
2852 pte_t *pte, entry;
2853 entry = pfn_pte(my_zero_pfn(haddr), vma->vm_page_prot);
2854 entry = pte_mkspecial(entry);
2855 pte = pte_offset_map(&_pmd, haddr);
2856 VM_BUG_ON(!pte_none(*pte));
2857 set_pte_at(mm, haddr, pte, entry);
2858 pte_unmap(pte);
2859 }
2860 smp_wmb(); /* make pte visible before pmd */
2861 pmd_populate(mm, pmd, pgtable);
2862 put_huge_zero_page();
2863}
2864
2865static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
2866 unsigned long haddr, bool freeze)
2867{
2868 struct mm_struct *mm = vma->vm_mm;
2869 struct page *page;
2870 pgtable_t pgtable;
2871 pmd_t _pmd;
2872 bool young, write, dirty;
2873 unsigned long addr;
2874 int i;
2875
2876 VM_BUG_ON(haddr & ~HPAGE_PMD_MASK);
2877 VM_BUG_ON_VMA(vma->vm_start > haddr, vma);
2878 VM_BUG_ON_VMA(vma->vm_end < haddr + HPAGE_PMD_SIZE, vma);
2879 VM_BUG_ON(!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd));
2880
2881 count_vm_event(THP_SPLIT_PMD);
2882
2883 if (vma_is_dax(vma)) {
2884 pmd_t _pmd = pmdp_huge_clear_flush_notify(vma, haddr, pmd);
2885 if (is_huge_zero_pmd(_pmd))
2886 put_huge_zero_page();
2887 return;
2888 } else if (is_huge_zero_pmd(*pmd)) {
2889 return __split_huge_zero_page_pmd(vma, haddr, pmd);
2890 }
2891
2892 page = pmd_page(*pmd);
2893 VM_BUG_ON_PAGE(!page_count(page), page);
2894 page_ref_add(page, HPAGE_PMD_NR - 1);
2895 write = pmd_write(*pmd);
2896 young = pmd_young(*pmd);
2897 dirty = pmd_dirty(*pmd);
2898
2899 pmdp_huge_split_prepare(vma, haddr, pmd);
2900 pgtable = pgtable_trans_huge_withdraw(mm, pmd);
2901 pmd_populate(mm, &_pmd, pgtable);
2902
2903 for (i = 0, addr = haddr; i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE) {
2904 pte_t entry, *pte;
2905 /*
2906 * Note that NUMA hinting access restrictions are not
2907 * transferred to avoid any possibility of altering
2908 * permissions across VMAs.
2909 */
2910 if (freeze) {
2911 swp_entry_t swp_entry;
2912 swp_entry = make_migration_entry(page + i, write);
2913 entry = swp_entry_to_pte(swp_entry);
2914 } else {
2915 entry = mk_pte(page + i, vma->vm_page_prot);
2916 entry = maybe_mkwrite(entry, vma);
2917 if (!write)
2918 entry = pte_wrprotect(entry);
2919 if (!young)
2920 entry = pte_mkold(entry);
2921 }
2922 if (dirty)
2923 SetPageDirty(page + i);
2924 pte = pte_offset_map(&_pmd, addr);
2925 BUG_ON(!pte_none(*pte));
2926 set_pte_at(mm, addr, pte, entry);
2927 atomic_inc(&page[i]._mapcount);
2928 pte_unmap(pte);
2929 }
2930
2931 /*
2932 * Set PG_double_map before dropping compound_mapcount to avoid
2933 * false-negative page_mapped().
2934 */
2935 if (compound_mapcount(page) > 1 && !TestSetPageDoubleMap(page)) {
2936 for (i = 0; i < HPAGE_PMD_NR; i++)
2937 atomic_inc(&page[i]._mapcount);
2938 }
2939
2940 if (atomic_add_negative(-1, compound_mapcount_ptr(page))) {
2941 /* Last compound_mapcount is gone. */
2942 __dec_zone_page_state(page, NR_ANON_TRANSPARENT_HUGEPAGES);
2943 if (TestClearPageDoubleMap(page)) {
2944 /* No need in mapcount reference anymore */
2945 for (i = 0; i < HPAGE_PMD_NR; i++)
2946 atomic_dec(&page[i]._mapcount);
2947 }
2948 }
2949
2950 smp_wmb(); /* make pte visible before pmd */
2951 /*
2952 * Up to this point the pmd is present and huge and userland has the
2953 * whole access to the hugepage during the split (which happens in
2954 * place). If we overwrite the pmd with the not-huge version pointing
2955 * to the pte here (which of course we could if all CPUs were bug
2956 * free), userland could trigger a small page size TLB miss on the
2957 * small sized TLB while the hugepage TLB entry is still established in
2958 * the huge TLB. Some CPU doesn't like that.
2959 * See http://support.amd.com/us/Processor_TechDocs/41322.pdf, Erratum
2960 * 383 on page 93. Intel should be safe but is also warns that it's
2961 * only safe if the permission and cache attributes of the two entries
2962 * loaded in the two TLB is identical (which should be the case here).
2963 * But it is generally safer to never allow small and huge TLB entries
2964 * for the same virtual address to be loaded simultaneously. So instead
2965 * of doing "pmd_populate(); flush_pmd_tlb_range();" we first mark the
2966 * current pmd notpresent (atomically because here the pmd_trans_huge
2967 * and pmd_trans_splitting must remain set at all times on the pmd
2968 * until the split is complete for this pmd), then we flush the SMP TLB
2969 * and finally we write the non-huge version of the pmd entry with
2970 * pmd_populate.
2971 */
2972 pmdp_invalidate(vma, haddr, pmd);
2973 pmd_populate(mm, pmd, pgtable);
2974
2975 if (freeze) {
2976 for (i = 0; i < HPAGE_PMD_NR; i++) {
2977 page_remove_rmap(page + i, false);
2978 put_page(page + i);
2979 }
2980 }
2981}
2982
2983void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
2984 unsigned long address, bool freeze)
2985{
2986 spinlock_t *ptl;
2987 struct mm_struct *mm = vma->vm_mm;
2988 unsigned long haddr = address & HPAGE_PMD_MASK;
2989
2990 mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
2991 ptl = pmd_lock(mm, pmd);
2992 if (pmd_trans_huge(*pmd)) {
2993 struct page *page = pmd_page(*pmd);
2994 if (PageMlocked(page))
2995 clear_page_mlock(page);
2996 } else if (!pmd_devmap(*pmd))
2997 goto out;
2998 __split_huge_pmd_locked(vma, pmd, haddr, freeze);
2999out:
3000 spin_unlock(ptl);
3001 mmu_notifier_invalidate_range_end(mm, haddr, haddr + HPAGE_PMD_SIZE);
3002}
3003
3004void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
3005 bool freeze, struct page *page)
3006{
3007 pgd_t *pgd;
3008 pud_t *pud;
3009 pmd_t *pmd;
3010
3011 pgd = pgd_offset(vma->vm_mm, address);
3012 if (!pgd_present(*pgd))
3013 return;
3014
3015 pud = pud_offset(pgd, address);
3016 if (!pud_present(*pud))
3017 return;
3018
3019 pmd = pmd_offset(pud, address);
3020 if (!pmd_present(*pmd) || (!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd)))
3021 return;
3022
3023 /*
3024 * If caller asks to setup a migration entries, we need a page to check
3025 * pmd against. Otherwise we can end up replacing wrong page.
3026 */
3027 VM_BUG_ON(freeze && !page);
3028 if (page && page != pmd_page(*pmd))
3029 return;
3030
3031 /*
3032 * Caller holds the mmap_sem write mode, so a huge pmd cannot
3033 * materialize from under us.
3034 */
3035 __split_huge_pmd(vma, pmd, address, freeze);
3036}
3037
3038void vma_adjust_trans_huge(struct vm_area_struct *vma,
3039 unsigned long start,
3040 unsigned long end,
3041 long adjust_next)
3042{
3043 /*
3044 * If the new start address isn't hpage aligned and it could
3045 * previously contain an hugepage: check if we need to split
3046 * an huge pmd.
3047 */
3048 if (start & ~HPAGE_PMD_MASK &&
3049 (start & HPAGE_PMD_MASK) >= vma->vm_start &&
3050 (start & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3051 split_huge_pmd_address(vma, start, false, NULL);
3052
3053 /*
3054 * If the new end address isn't hpage aligned and it could
3055 * previously contain an hugepage: check if we need to split
3056 * an huge pmd.
3057 */
3058 if (end & ~HPAGE_PMD_MASK &&
3059 (end & HPAGE_PMD_MASK) >= vma->vm_start &&
3060 (end & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= vma->vm_end)
3061 split_huge_pmd_address(vma, end, false, NULL);
3062
3063 /*
3064 * If we're also updating the vma->vm_next->vm_start, if the new
3065 * vm_next->vm_start isn't page aligned and it could previously
3066 * contain an hugepage: check if we need to split an huge pmd.
3067 */
3068 if (adjust_next > 0) {
3069 struct vm_area_struct *next = vma->vm_next;
3070 unsigned long nstart = next->vm_start;
3071 nstart += adjust_next << PAGE_SHIFT;
3072 if (nstart & ~HPAGE_PMD_MASK &&
3073 (nstart & HPAGE_PMD_MASK) >= next->vm_start &&
3074 (nstart & HPAGE_PMD_MASK) + HPAGE_PMD_SIZE <= next->vm_end)
3075 split_huge_pmd_address(next, nstart, false, NULL);
3076 }
3077}
3078
3079static void freeze_page(struct page *page)
3080{
3081 enum ttu_flags ttu_flags = TTU_MIGRATION | TTU_IGNORE_MLOCK |
3082 TTU_IGNORE_ACCESS | TTU_RMAP_LOCKED;
3083 int i, ret;
3084
3085 VM_BUG_ON_PAGE(!PageHead(page), page);
3086
3087 /* We only need TTU_SPLIT_HUGE_PMD once */
3088 ret = try_to_unmap(page, ttu_flags | TTU_SPLIT_HUGE_PMD);
3089 for (i = 1; !ret && i < HPAGE_PMD_NR; i++) {
3090 /* Cut short if the page is unmapped */
3091 if (page_count(page) == 1)
3092 return;
3093
3094 ret = try_to_unmap(page + i, ttu_flags);
3095 }
3096 VM_BUG_ON(ret);
3097}
3098
3099static void unfreeze_page(struct page *page)
3100{
3101 int i;
3102
3103 for (i = 0; i < HPAGE_PMD_NR; i++)
3104 remove_migration_ptes(page + i, page + i, true);
3105}
3106
3107static void __split_huge_page_tail(struct page *head, int tail,
3108 struct lruvec *lruvec, struct list_head *list)
3109{
3110 struct page *page_tail = head + tail;
3111
3112 VM_BUG_ON_PAGE(atomic_read(&page_tail->_mapcount) != -1, page_tail);
3113 VM_BUG_ON_PAGE(page_ref_count(page_tail) != 0, page_tail);
3114
3115 /*
3116 * tail_page->_count is zero and not changing from under us. But
3117 * get_page_unless_zero() may be running from under us on the
3118 * tail_page. If we used atomic_set() below instead of atomic_inc(), we
3119 * would then run atomic_set() concurrently with
3120 * get_page_unless_zero(), and atomic_set() is implemented in C not
3121 * using locked ops. spin_unlock on x86 sometime uses locked ops
3122 * because of PPro errata 66, 92, so unless somebody can guarantee
3123 * atomic_set() here would be safe on all archs (and not only on x86),
3124 * it's safer to use atomic_inc().
3125 */
3126 page_ref_inc(page_tail);
3127
3128 page_tail->flags &= ~PAGE_FLAGS_CHECK_AT_PREP;
3129 page_tail->flags |= (head->flags &
3130 ((1L << PG_referenced) |
3131 (1L << PG_swapbacked) |
3132 (1L << PG_mlocked) |
3133 (1L << PG_uptodate) |
3134 (1L << PG_active) |
3135 (1L << PG_locked) |
3136 (1L << PG_unevictable) |
3137 (1L << PG_dirty)));
3138
3139 /*
3140 * After clearing PageTail the gup refcount can be released.
3141 * Page flags also must be visible before we make the page non-compound.
3142 */
3143 smp_wmb();
3144
3145 clear_compound_head(page_tail);
3146
3147 if (page_is_young(head))
3148 set_page_young(page_tail);
3149 if (page_is_idle(head))
3150 set_page_idle(page_tail);
3151
3152 /* ->mapping in first tail page is compound_mapcount */
3153 VM_BUG_ON_PAGE(tail > 2 && page_tail->mapping != TAIL_MAPPING,
3154 page_tail);
3155 page_tail->mapping = head->mapping;
3156
3157 page_tail->index = head->index + tail;
3158 page_cpupid_xchg_last(page_tail, page_cpupid_last(head));
3159 lru_add_page_tail(head, page_tail, lruvec, list);
3160}
3161
3162static void __split_huge_page(struct page *page, struct list_head *list)
3163{
3164 struct page *head = compound_head(page);
3165 struct zone *zone = page_zone(head);
3166 struct lruvec *lruvec;
3167 int i;
3168
3169 /* prevent PageLRU to go away from under us, and freeze lru stats */
3170 spin_lock_irq(&zone->lru_lock);
3171 lruvec = mem_cgroup_page_lruvec(head, zone);
3172
3173 /* complete memcg works before add pages to LRU */
3174 mem_cgroup_split_huge_fixup(head);
3175
3176 for (i = HPAGE_PMD_NR - 1; i >= 1; i--)
3177 __split_huge_page_tail(head, i, lruvec, list);
3178
3179 ClearPageCompound(head);
3180 spin_unlock_irq(&zone->lru_lock);
3181
3182 unfreeze_page(head);
3183
3184 for (i = 0; i < HPAGE_PMD_NR; i++) {
3185 struct page *subpage = head + i;
3186 if (subpage == page)
3187 continue;
3188 unlock_page(subpage);
3189
3190 /*
3191 * Subpages may be freed if there wasn't any mapping
3192 * like if add_to_swap() is running on a lru page that
3193 * had its mapping zapped. And freeing these pages
3194 * requires taking the lru_lock so we do the put_page
3195 * of the tail pages after the split is complete.
3196 */
3197 put_page(subpage);
3198 }
3199}
3200
3201int total_mapcount(struct page *page)
3202{
3203 int i, ret;
3204
3205 VM_BUG_ON_PAGE(PageTail(page), page);
3206
3207 if (likely(!PageCompound(page)))
3208 return atomic_read(&page->_mapcount) + 1;
3209
3210 ret = compound_mapcount(page);
3211 if (PageHuge(page))
3212 return ret;
3213 for (i = 0; i < HPAGE_PMD_NR; i++)
3214 ret += atomic_read(&page[i]._mapcount) + 1;
3215 if (PageDoubleMap(page))
3216 ret -= HPAGE_PMD_NR;
3217 return ret;
3218}
3219
3220/*
3221 * This calculates accurately how many mappings a transparent hugepage
3222 * has (unlike page_mapcount() which isn't fully accurate). This full
3223 * accuracy is primarily needed to know if copy-on-write faults can
3224 * reuse the page and change the mapping to read-write instead of
3225 * copying them. At the same time this returns the total_mapcount too.
3226 *
3227 * The function returns the highest mapcount any one of the subpages
3228 * has. If the return value is one, even if different processes are
3229 * mapping different subpages of the transparent hugepage, they can
3230 * all reuse it, because each process is reusing a different subpage.
3231 *
3232 * The total_mapcount is instead counting all virtual mappings of the
3233 * subpages. If the total_mapcount is equal to "one", it tells the
3234 * caller all mappings belong to the same "mm" and in turn the
3235 * anon_vma of the transparent hugepage can become the vma->anon_vma
3236 * local one as no other process may be mapping any of the subpages.
3237 *
3238 * It would be more accurate to replace page_mapcount() with
3239 * page_trans_huge_mapcount(), however we only use
3240 * page_trans_huge_mapcount() in the copy-on-write faults where we
3241 * need full accuracy to avoid breaking page pinning, because
3242 * page_trans_huge_mapcount() is slower than page_mapcount().
3243 */
3244int page_trans_huge_mapcount(struct page *page, int *total_mapcount)
3245{
3246 int i, ret, _total_mapcount, mapcount;
3247
3248 /* hugetlbfs shouldn't call it */
3249 VM_BUG_ON_PAGE(PageHuge(page), page);
3250
3251 if (likely(!PageTransCompound(page))) {
3252 mapcount = atomic_read(&page->_mapcount) + 1;
3253 if (total_mapcount)
3254 *total_mapcount = mapcount;
3255 return mapcount;
3256 }
3257
3258 page = compound_head(page);
3259
3260 _total_mapcount = ret = 0;
3261 for (i = 0; i < HPAGE_PMD_NR; i++) {
3262 mapcount = atomic_read(&page[i]._mapcount) + 1;
3263 ret = max(ret, mapcount);
3264 _total_mapcount += mapcount;
3265 }
3266 if (PageDoubleMap(page)) {
3267 ret -= 1;
3268 _total_mapcount -= HPAGE_PMD_NR;
3269 }
3270 mapcount = compound_mapcount(page);
3271 ret += mapcount;
3272 _total_mapcount += mapcount;
3273 if (total_mapcount)
3274 *total_mapcount = _total_mapcount;
3275 return ret;
3276}
3277
3278/*
3279 * This function splits huge page into normal pages. @page can point to any
3280 * subpage of huge page to split. Split doesn't change the position of @page.
3281 *
3282 * Only caller must hold pin on the @page, otherwise split fails with -EBUSY.
3283 * The huge page must be locked.
3284 *
3285 * If @list is null, tail pages will be added to LRU list, otherwise, to @list.
3286 *
3287 * Both head page and tail pages will inherit mapping, flags, and so on from
3288 * the hugepage.
3289 *
3290 * GUP pin and PG_locked transferred to @page. Rest subpages can be freed if
3291 * they are not mapped.
3292 *
3293 * Returns 0 if the hugepage is split successfully.
3294 * Returns -EBUSY if the page is pinned or if anon_vma disappeared from under
3295 * us.
3296 */
3297int split_huge_page_to_list(struct page *page, struct list_head *list)
3298{
3299 struct page *head = compound_head(page);
3300 struct pglist_data *pgdata = NODE_DATA(page_to_nid(head));
3301 struct anon_vma *anon_vma;
3302 int count, mapcount, ret;
3303 bool mlocked;
3304 unsigned long flags;
3305
3306 VM_BUG_ON_PAGE(is_huge_zero_page(page), page);
3307 VM_BUG_ON_PAGE(!PageAnon(page), page);
3308 VM_BUG_ON_PAGE(!PageLocked(page), page);
3309 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
3310 VM_BUG_ON_PAGE(!PageCompound(page), page);
3311
3312 /*
3313 * The caller does not necessarily hold an mmap_sem that would prevent
3314 * the anon_vma disappearing so we first we take a reference to it
3315 * and then lock the anon_vma for write. This is similar to
3316 * page_lock_anon_vma_read except the write lock is taken to serialise
3317 * against parallel split or collapse operations.
3318 */
3319 anon_vma = page_get_anon_vma(head);
3320 if (!anon_vma) {
3321 ret = -EBUSY;
3322 goto out;
3323 }
3324 anon_vma_lock_write(anon_vma);
3325
3326 /*
3327 * Racy check if we can split the page, before freeze_page() will
3328 * split PMDs
3329 */
3330 if (total_mapcount(head) != page_count(head) - 1) {
3331 ret = -EBUSY;
3332 goto out_unlock;
3333 }
3334
3335 mlocked = PageMlocked(page);
3336 freeze_page(head);
3337 VM_BUG_ON_PAGE(compound_mapcount(head), head);
3338
3339 /* Make sure the page is not on per-CPU pagevec as it takes pin */
3340 if (mlocked)
3341 lru_add_drain();
3342
3343 /* Prevent deferred_split_scan() touching ->_count */
3344 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3345 count = page_count(head);
3346 mapcount = total_mapcount(head);
3347 if (!mapcount && count == 1) {
3348 if (!list_empty(page_deferred_list(head))) {
3349 pgdata->split_queue_len--;
3350 list_del(page_deferred_list(head));
3351 }
3352 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3353 __split_huge_page(page, list);
3354 ret = 0;
3355 } else if (IS_ENABLED(CONFIG_DEBUG_VM) && mapcount) {
3356 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3357 pr_alert("total_mapcount: %u, page_count(): %u\n",
3358 mapcount, count);
3359 if (PageTail(page))
3360 dump_page(head, NULL);
3361 dump_page(page, "total_mapcount(head) > 0");
3362 BUG();
3363 } else {
3364 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3365 unfreeze_page(head);
3366 ret = -EBUSY;
3367 }
3368
3369out_unlock:
3370 anon_vma_unlock_write(anon_vma);
3371 put_anon_vma(anon_vma);
3372out:
3373 count_vm_event(!ret ? THP_SPLIT_PAGE : THP_SPLIT_PAGE_FAILED);
3374 return ret;
3375}
3376
3377void free_transhuge_page(struct page *page)
3378{
3379 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
3380 unsigned long flags;
3381
3382 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3383 if (!list_empty(page_deferred_list(page))) {
3384 pgdata->split_queue_len--;
3385 list_del(page_deferred_list(page));
3386 }
3387 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3388 free_compound_page(page);
3389}
3390
3391void deferred_split_huge_page(struct page *page)
3392{
3393 struct pglist_data *pgdata = NODE_DATA(page_to_nid(page));
3394 unsigned long flags;
3395
3396 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3397
3398 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3399 if (list_empty(page_deferred_list(page))) {
3400 count_vm_event(THP_DEFERRED_SPLIT_PAGE);
3401 list_add_tail(page_deferred_list(page), &pgdata->split_queue);
3402 pgdata->split_queue_len++;
3403 }
3404 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3405}
3406
3407static unsigned long deferred_split_count(struct shrinker *shrink,
3408 struct shrink_control *sc)
3409{
3410 struct pglist_data *pgdata = NODE_DATA(sc->nid);
3411 return ACCESS_ONCE(pgdata->split_queue_len);
3412}
3413
3414static unsigned long deferred_split_scan(struct shrinker *shrink,
3415 struct shrink_control *sc)
3416{
3417 struct pglist_data *pgdata = NODE_DATA(sc->nid);
3418 unsigned long flags;
3419 LIST_HEAD(list), *pos, *next;
3420 struct page *page;
3421 int split = 0;
3422
3423 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3424 /* Take pin on all head pages to avoid freeing them under us */
3425 list_for_each_safe(pos, next, &pgdata->split_queue) {
3426 page = list_entry((void *)pos, struct page, mapping);
3427 page = compound_head(page);
3428 if (get_page_unless_zero(page)) {
3429 list_move(page_deferred_list(page), &list);
3430 } else {
3431 /* We lost race with put_compound_page() */
3432 list_del_init(page_deferred_list(page));
3433 pgdata->split_queue_len--;
3434 }
3435 if (!--sc->nr_to_scan)
3436 break;
3437 }
3438 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3439
3440 list_for_each_safe(pos, next, &list) {
3441 page = list_entry((void *)pos, struct page, mapping);
3442 lock_page(page);
3443 /* split_huge_page() removes page from list on success */
3444 if (!split_huge_page(page))
3445 split++;
3446 unlock_page(page);
3447 put_page(page);
3448 }
3449
3450 spin_lock_irqsave(&pgdata->split_queue_lock, flags);
3451 list_splice_tail(&list, &pgdata->split_queue);
3452 spin_unlock_irqrestore(&pgdata->split_queue_lock, flags);
3453
3454 /*
3455 * Stop shrinker if we didn't split any page, but the queue is empty.
3456 * This can happen if pages were freed under us.
3457 */
3458 if (!split && list_empty(&pgdata->split_queue))
3459 return SHRINK_STOP;
3460 return split;
3461}
3462
3463static struct shrinker deferred_split_shrinker = {
3464 .count_objects = deferred_split_count,
3465 .scan_objects = deferred_split_scan,
3466 .seeks = DEFAULT_SEEKS,
3467 .flags = SHRINKER_NUMA_AWARE,
3468};
3469
3470#ifdef CONFIG_DEBUG_FS
3471static int split_huge_pages_set(void *data, u64 val)
3472{
3473 struct zone *zone;
3474 struct page *page;
3475 unsigned long pfn, max_zone_pfn;
3476 unsigned long total = 0, split = 0;
3477
3478 if (val != 1)
3479 return -EINVAL;
3480
3481 for_each_populated_zone(zone) {
3482 max_zone_pfn = zone_end_pfn(zone);
3483 for (pfn = zone->zone_start_pfn; pfn < max_zone_pfn; pfn++) {
3484 if (!pfn_valid(pfn))
3485 continue;
3486
3487 page = pfn_to_page(pfn);
3488 if (!get_page_unless_zero(page))
3489 continue;
3490
3491 if (zone != page_zone(page))
3492 goto next;
3493
3494 if (!PageHead(page) || !PageAnon(page) ||
3495 PageHuge(page))
3496 goto next;
3497
3498 total++;
3499 lock_page(page);
3500 if (!split_huge_page(page))
3501 split++;
3502 unlock_page(page);
3503next:
3504 put_page(page);
3505 }
3506 }
3507
3508 pr_info("%lu of %lu THP split\n", split, total);
3509
3510 return 0;
3511}
3512DEFINE_SIMPLE_ATTRIBUTE(split_huge_pages_fops, NULL, split_huge_pages_set,
3513 "%llu\n");
3514
3515static int __init split_huge_pages_debugfs(void)
3516{
3517 void *ret;
3518
3519 ret = debugfs_create_file("split_huge_pages", 0200, NULL, NULL,
3520 &split_huge_pages_fops);
3521 if (!ret)
3522 pr_warn("Failed to create split_huge_pages in debugfs");
3523 return 0;
3524}
3525late_initcall(split_huge_pages_debugfs);
3526#endif